论文信息

Jassim M. A. Al-A'arej / Zaid J. Al-Araji / Fahad Ahmed Shaban 等 7 人

Journal of Applied Science and Technology Trends

论文解读

AI 辅助整理
方法
EfficientNet-B0全局分类器、多边形监督ROI网络与局部分类器,置信
发现
平均测试准确率98.78%,但ROI引导相较全局分类器的提升无统计显著性。
机会点
可探索轻量化田间部署与独立田间数据泛化验证,并量化ROI引导的真实增益。

论文摘要

由于病害类别之间视觉相似、症状表现存在差异以及类别不平衡,准确的草莓病害分类仍具挑战性。传统的整图分类器可能忽略可补充全局视觉表示的局部病害相关信息。本研究提出StrawNet-ROI,一种标注引导的双视图深度学习框架,集成了多尺度EfficientNet-B0全局分类器、多边形监督的兴趣区域(ROI)预测网络以及单独微调的局部分类器。基于连通域的过程从预测的ROI图中生成空间受限的裁剪区域,而置信度自适应概率融合在推理阶段无需真实标注即可结合全局与局部预测。实验在一个经审核的七类草莓病害数据集上进行,该数据集包含2,457张图像,划分为1,426个训练样本、294个验证样本和737个测试样本。在五个随机种子下,StrawNet-ROI取得了平均测试准确率98.78 ± 0.25%、宏精确率98.26 ± 0.60%、宏召回率97.40 ± 0.43%和宏F1值97.79 ± 0.47%。与仅使用全局分类器相比,该框架在宏F1值上取得了0.59个百分点的数值提升;然而,配对差异在统计上不显著(p = 0.1268)。额外的裁剪对照实验未能确立ROI引导选择在统计上的显著优越性。完整框架包含1,258万个参数,在服务器级GPU上实现了18.80 ms的平均端到端推理延迟。这些发现表明,全局–局部表示融合可能为草莓病害分类提供互补信息,但额外的计算成本和有限的统计证据值得谨慎对待。由于评估依赖于历史审核的内部测试划分,独立的田间泛化能力和资源受限部署仍未经验证。

全文

Journal of Applied Science and Technology Trends

Vol. 07, No. 02, pp. 534-553 (2026)
ISSN: 2708-0757

JOURNAL OF APPLIED SCIENCE AND TECHNOLOGY TRENDS

www.jastt.org

534
doi: https://doi.org/10.38094/jastt721094
StrawNet-ROI: An Annotation-Guided Dual-View Framework
for Strawberry Disease Classification

Jassim M. A. Al-A'areji1, Zaid J. Al-Araji2*, Fahad Ahmed Shaban2, Balqees Talal Hasan3, Hussam Ahmed Rammo4, Ahmed A.
Idris5, Nawar Talal Thanoon2
1 Horticulture and Landscape Design Department, College of Agriculture and Forestry, Mosul University, Iraq
2Department of Computer Networks and Internet, College of Information Technology, Ninevah University, Mosul, Iraq
3Department of Computer Artificial Intelligence, College of Information Technology, Ninevah University, Mosul, Iraq
4Techniques Engineering College for Computer and AI, Northern Technical University, Mosul, Iraq
5Department of Computer Science, Al-Hamdaniya University, Mosul, Iraq
jassim@uomosul.edu.iq , zaid.jasim@uoninevah.edu.iq; fahad.ahmed@uoninevah.edu.iq; nawar.talal@uoninevah.edu.iq
balqees.hasan@uoninevah.edu.iq ,hussam.rammo@ntu.edu.iq ,ahmed-alkaddo@uohamdaniya.edu.iq
Correspondence: zaid.jasim@uoninevah.edu.iq

Abstract
Accurate strawberry disease classification remains challenging due to visual similarities among disease categories, variations in symptom
appearance, and class imbalance. Conventional whole-image classifiers may overlook localized disease-associated information that could
complement global visual representations. This study proposes StrawNet-ROI, an annotation-guided dual-view deep learning framework
integrating a multi -scale EfficientNet -B0 global classifier, a polygon -supervised region -of-interest (ROI) prediction network, and a
separately fine-tuned local classifier. A connected-component-based procedure generates spatially constrained crops from predicted ROI
maps, while confidence -adaptive probability fusion combines global and local predictions without requiring ground -truth annotations
during inference. Experiments were conducted on an audited seven-class strawberry disease dataset comprising 2,457 images, partitioned
into 1,426 training, 294 validation, and 737 testing samples. Across five random seeds, StrawNet-ROI achieved a mean test accuracy of
98.78 ± 0.25%, macro -precision of 98.26 ± 0.60%, macro -recall of 97.40 ± 0.43%, and macro -F1 of 97.79 ± 0.47%. Compared with the
global-only classifier, the framework achieved a numerical macro -F1 improvement of 0.59 p ercentage points; however, the paired
difference was not statistically significant (p = 0.1268). Additional cropping controls did not establish statistically signi ficant superiority
of ROI-guided selection. The complete framework contains 12.58 million par ameters and achieved a mean end -to-end inference latency
of 18.80 ms on a server-class GPU. These findings suggest that global–local representation fusion may provide complementary information
for strawberry disease classification, although the additional computational cost and limited statistical evidence warrant caution. Because
the evaluation relied on a historically examined internal test partition, independent field generalization and resource -constrained
deployment remain unverified.

Keywords: Strawberry leaf disease, lightweight deep learning, smart agriculture, CNN
Received: April 20th, 2026 / Revised: October 2nd, 2026 / Accepted: October 28th, 2026 / Online: November 16th, 2026

I. INTRODUCTION
Timely detection of plant diseases is fundamental in disease
management as any infection is expected to negatively influence
crop yield, fruit quality and the success of future interventions
[1], [2]. There is increased vulnerability in strawberry cultivation
for visual estimation of disease. Symptoms can be found on
leaves, flowers and fruits and they may change in incubation
time, lighting environment and tissue status [3], [4] . Some
diseases share common symptoms, like discoloration, necrotic
spots and surface growth. In RGB images acquired by
conventional cameras, these symptoms may only represent a
small part of the image or be covered by other plants. Therefore,
a robust classifier for the data is presumed to need detecting
small patterns in the image as well as the large variation of visual
appearances in the picture . Plant disease recognition probably
also benefits from a methodological standardization based on

Al-A'areji et al. / Journal of Applied Science and Technology Trends Vol. 07, No. 02, pp. 534-553, (2026)

535
deep learning. Convolutional neural networks could learn local
discriminatory texture and shape representations from labeled
images; pretrained architectures could confirm that these
representations can be transferred to a small number of
agricultural samples [5], [6], [7] . However, being highly
predictive on a curated internal holdout set does not guarantee
similar performances on novel production conditions. Variations
in acquisition devices, background information, disease
prevalence and growing conditions may generate dis tribution
shift that are unlikely to be captured by standard random splits
[8], [9], [10]. For strawberry disease classification, this concern
is compounded by differences in class taxonomy and evaluation
protocols across published datasets, which complicate
comparisons between methods [4], [11], [12].
Recent strawberry -specific studies have pursued several
directions. BerryNet-Lite [13] and Light-Mobile Berry Net [14]
emphasize efficient image classification, whereas ENet-CAEM
[11] and the SCSA -Transformer [15] investigate contextual or
attention-based representation learning. Other approaches
formulate the problem as object detection or instance
segmentation, thereby locating disease-associated plant regions
rather than assigning only an image -level label [3], [4], [16],
[17]. These contributions are complementary but cannot be
placed on a common performance scale without accounting for
different disease categories, datasets, test partitions, prediction
targets, and computational measurement procedures. In
particular, mean average precision for a detector should not be
interpreted as equivalent to macro -F1 for a seven -class image
classifier.
This study was prompted by two ideas of methods. Whole-
image classification, though expected to keep context, tends to
allocate representational capacity among larger regions of
healthy tissue and background. Cropping may accentuate local
symptoms, but a l ocal view risks neglecting organ -scene
information at higher resolution scales. In addition, common
feature-pyramid and attention modules imply an aggregation of
information through resolutions or a rescaling of intermediate
feature activations [18], [19], [20] ; These results do not
definitively show that an automatically selected image region
can give complementary information to a global classifier. And
adding additional localization and classification modules adds
computational expense that might offset the sim plicity
advantage of a lean backbone. It is important to consider local
selection and prediction fusion together with inference
complexity, rather than simply infer from a given architecture an
improved accuracy or suitability for deployment.
The literature reported underscores the importance of
carefully handling data imbalance and evaluation validity. As
minority disease categories are expected to contribute less to
overall accuracy, it may call for macro -averaged measures and
frequency-aware learning objectives [21], [22], [23] .
Nevertheless, it is crucial to disentangle the impact of class
weighting, focal modulation, and the individual attentional
modules from the global system efficiency using suitable
experiments. Likewise, duplicates, or visually related samples
over partitions may overestimate the internal validation. While a
grouping based on similarity can prevent this, the independence
of biological samples or acquisition sessions cannot be garanteed
in the absence of provenance [8], [9] . These considerations
define the scope of a defensible evaluation.
Against this background, the present study examines
whether predicted disease-associated regions can provide useful
local evidence alongside full-image representations. We propose
Straw
Net-ROI, an annotation -guided dual -view framework for
seven-class strawberry disease classification. The system
combines an ImageNet -pretrained EfficientNet -B0-based
multiscale global classifier, a polygon -supervised region -of-
interest (ROI) predictor, a d eterministic connected-component-
guided crop -selection procedure, and a separately fine -tuned
local classifier. Global and local class probabilities are
integrated using a validation -selected confidence -adaptive
fusion rule. Polygon masks supervise ROI lea rning during
training only; test-time crops are generated from predicted maps
without access to ground-truth spatial annotations. Because the
source polygons sometimes encompass affected plant organs
rather than exact lesions, the resulting ROI maps are no t
presented as pixel-accurate lesion segmentations.
The principal contributions of this work are as follows:
(1) An annotation -guided dual -view architecture that
integrates global classification, predicted ROI localization,
spatially constrained local cropping, and independent local
classification without requiring test-time polygon annotations.
(2) A controlled investigation of local -view selection and
decision fusion, including global-only, ROI-only, fixed-weight
and confidence-adaptive fusion, and size -matched center- and
random-crop alternatives.
(3) A reproducible exploratory evaluation on a duplicate -
audited seven-class strawberry dataset using five random seeds,
class-wise metrics, architectural sensitivity analyses, and
explicit acknowledgement of data -partition and statistical
limitations.
(4) A transparent assessment of the predictive –
computational trade-off using model-parameter accounting and
end-to-end inference measurements, without asserting mobile -
device or field -deployment capability in the absence of
corresponding experiments.
The remainder of this paper is organized as follows. Section
II critically reviews related work on strawberry disease
classification, spatial localization, class imbalance, and
computational efficiency. Section III describes the proposed
StrawNet-ROI architecture and training procedure. Section IV
presents the dataset audit, experimental protocol, comparative
evaluation, ablations, and computational measurements. Section
V concludes the paper, discusses limitations, and identifies
directions for i ndependent validation and deployment -oriented
research.

Al-A'areji et al. / Journal of Applied Science and Technology Trends Vol. 07, No. 02, pp. 534-553, (2026)

536
II. RELATED WORK
Initially, image-based plant disease recognition was driven
by convolutional general classification toward handling
acquisition variability, localizing space, and meeting operational
needs. Early work has shown that deep networks can recognize
many plant disease classes in curated image repositories [5], [6]
and transfer learning pointed to pretrained representations for
small-to-moderate agricultural training sets [7], enabling fast
architectural testing but exposing a major gap in operational
evidence. Barbedo [8] hypothesised factors which would
influence plant diagnosis by deep learning, such as image quality
and data composition; as well as symptom interpretation
challenges under less -standardised image acquisition, such as
that tested by the Plant Doc [9] and Plant Pathology Challenge
dataset [23]. Similar reviews of agricultural imaging call for
matching the prediction task, sensing condition, and validation
design with the use case [2], [24].
In strawberry disease classification, transfer learning,
receptive-field expansion and efficient channel attention are
adopted by Berry Net -Lite [13]. The result may be used as a
guidance to model design, but not easy to be comparable with
the seven -class classification trained on the Afzaal image
collection. Li et al [15] proposed SCSA-Transformer to learn
spatial and contextual information for complex scenes. The
authors created a data set of 5,369 images of 12 classes of
strawberry diseases. This work demonstrates the relevance of
richer feature interactions, although its different class taxonomy
and experimental data prevent a like-for-like numerical ranking
against StrawNet-ROI.
ENet-CAEM [11], which is based on Efficient Net -B0,
channel-context modeling, multiscale attention, and lightweight
spatial-pyramid computation strategies. Chang et al. estimate
85.84% accuracy using a strawberry source dataset, and 97.39%
on a public dataset of compara ble size with seven disease
classes.
Curiously, although their public dataset appears to cover
the same seven classes as the source dataset for this paper, their
pre-processing, splits, and training design differ from this
proposed audited testing, while the table of their fp -ts reports
0.461 GFLOPs, but the use of an ambiguous unit label in the
parameter column complicates reader attempts to store such data
without standard measurement procedures. ENet -CAEM is
therefore probably only relevant as context for multiscale
classification rather than as a template for an exact experiment.
Light-MobileBerry Net [14] offers an important
complementary comparison. Based on an efficient
MobileNetV3-inspired design, it uses the public Strawberry
Disease Detection Dataset and reports 96.60% classification
accuracy with approximately 0.53 million parameters and a
reported model footprint of about 2 MB. Unlike StrawNet-ROI,
its emphasis is on a single compact classification network and
qualitative Grad-CAM interpretation [25]. The use of a common
source collection makes the work especially relevant, but a
common source is not equivalent to an identical audited subset,
partition, or evaluation protocol. Its published accuracy and
model-size figures therefore illustrate a different predictive –
computational trade-off rather than establishing a direct head-to-
head comparison.
Other classification investigations address challenges that
are not captured by parameter count alone. Lee et al. [12]
examined an in -backbone latent -diffusion framework under
class imbalance and considered separately collected greenhouse
imagery for a four -class strawberry task. Such external testing
directly addresses acquisition shift, but its disease taxonomy and
training formulation differ from those in the present study. More
generally, long-tailed recognition research has investigated focal
modulation [21], effective-number-based class balancing [22],
targeted supervised contrastive learning [26], and subclass -
balanced representation learning [27]. These approaches make
different assumptions about how rare classes should be
emphasized. The proposed global classifier adopts an inverse -
square-root frequency -weighted focal objective, whereas its
local classifier uses class -weighted cross -entropy with label
smoothing. Neither objective should be conflated with an
effective-number loss, and the independent performance benefit
of either weighting mechanism is not asserted without a matched
ablation.
Localization-oriented strawberry studies provide a second,
distinct body of evidence. Afzaal et al. [4] introduced the
polygon-annotated seven -disease dataset used as the source
collection in the present work and evaluated an instance -
segmentation system based on Mask R -CNN. Zhao et al. [3]
combined ResNet, a feature pyramid, and attention in a Faster
R-CNN detector for diseases affecting strawberry leaves,
flowers, and fruits, reporting 92.18% mAP and an average
detection time of 229 ms under their study conditions. He et al.
[10] modified YOLOv8 for strawberry leaf disease detection,
while Amanova et al. [17] developed a two -stage
YOLOv10n/MobileViT-S system for detecting strawberry leaf
diseases and estimating three ordinal levels of leaf-spot severity
on embedded hardware. These methods underscore the
usefulness of spatially explicit predictions but perform detection,
segmentation, or severity estimation; their reported mAP or
ordinal-classification results must not be compared directly with
seven-class image-level macro-F1.
The methodological components used in many of these
studies are well established beyond agriculture. Feature Pyramid
Networks [18] combine hierarchical representations; channel -
recalibration methods such as squeeze -and-excitation [28] and
efficient channel attention [29], together with convolutional
block attention [19], modulate feature responses. Compact
mobile architectures [30] and lightweight plant -disease
classifiers [31], [32] seek to reduce the cost of inference.
Knowledge-distillation research [33], [36] including teacher -
assistant learning and plant-disease-specific student networks
[34], [35] , [37] , further demonstrates that predictive
performance and computational footprint can be optimized
through different training strategies. These works provide
appropriate conceptual context, but individual modules cannot
be considered original contributions me rely because they are
included in a new pipeline. Moreover, the final StrawNet -ROI

Al-A'areji et al. / Journal of Applied Science and Technology Trends Vol. 07, No. 02, pp. 534-553, (2026)

537
system does not use knowledge distillation, pruning, or
quantization.
A key distinction concerns what an ROI pathway actually
contributes. Post hoc heat maps, including Grad-CAM [25], can
visualize class -discriminative image regions but do not
constitute validated pixel -level lesion annotations. Likewise,
polygon supervision in the Afzaal dataset [4] does not
necessarily delineate the precise diseased tissue in every image.
In StrawNet -ROI, annotations supervise a region -prediction
network during training; the predicted probability map is
converted into a crop through deterministic postprocessing, and
a separately optimized classifier processes that crop at inference.
This arrangement should be evaluated against whole -image
classification, alternative cropping policies, and fixed fusion
before attributing any improvement to disease -specific region
selection. The most directly related studies and the limits of their
comparability are summarized in Table I.
TABLE I. COMPARATIVE ASSESSMENT OF RELATED STRAWBERRY DISEASE RECOGNITION STUDIES
Study / year Dataset and prediction scope Model size / computation Reported
performance
Evaluation distinction
[4] 2,500-source-image collection;
7 diseases; instance
segmentation
Mask R-CNN; comparable
inference cost not established
mAP 82.43% Source annotations;
segmentation metric is not
image-classification F1
[3] Greenhouse strawberry images;
disease object detection on plant
organs
ResNet/FPN/CBAM detector; 229
ms per image (reported)
mAP 92.18% Different collection and
detection target;
hardware/timing not matched
[11] Self-collected set plus public
2,500-image, 7-class set
0.461 GFLOPs; parameter-unit
label unclear in source
Accuracy 85.84%
(self-collected);
97.39% (public)
No shared audited split or equal
training budget
[14] Afzaal public seven-class
collection; image classification
≈0.53 M parameters; ≈2 MB
reported footprint
Accuracy 96.60% Same source collection, different
preprocessing and partition
[12] 4-class strawberry classification
with separate greenhouse
imagery
Matched model size and latency
not established
Field macro-F1
90.00% (reported)
Different disease labels; external
field testing
[17] 7-class leaf detection plus 3-
level leaf-spot severity
Two-stage detector/ordinal
classifier; Jetson Nano evaluation
Detection mAP@0.5
96.0%; severity
accuracy 97.3%
Different target and metrics;
edge hardware evaluated
Straw Net-ROI
(this study)
2,457 audited images; 7-class
image classification
12.58 M parameters; 50.31 MB
theoretical FP32; 18.80 ms server-
GPU end-to-end
Accuracy 98.78 ±
0.25%; macro-F1
97.79 ± 0.47%
5-seed internal test; previously
examined test data; no external
validation
The literature reviewed hints at three unresolved issues
relevant to the task at hand. First, strawberry classifiers in the
literature have tended towards whole -image representations or
architectural changes, while detection and segmentation
research is making a different prediction on the spatial problem
at hand. There is little controlled evidence among the studies
reviewed that a local view predicted automatically from an
annotation supervision adds information about the image -level
classifier that ordin ary crop does not. Second, improvements
reported are difficult to interpret across studies, given
differences in class coverage, dataset splits, pretrained
initialization, optimization budgets, and metrics; source-dataset
overlap alone is likely insufficie nt for a shared benchmark .
Third, using systems with localization and secondary classifier
often force to add computation cost, which is often not well
separated from parameters or forward -only latency for single -
backbone systems.
This work attempts to close this methodological gap with
the combination of supervised ROI prediction and global -local
classification. It compares a diverse set of systems including
ROI-only, center -crop, random -crop, fixed -fusion, and
confidence-adaptive systems on a systematically audited
common partition, with repeated -run metrics and an expanded
set of inference cost, always explicitly acknowledging its
statistical evidence as exploratory. With this scope, the positive
results of this approach on this set of benchmarks do not
empirically show the supremacy of each individual element,
consistency across tasks, or suitability for deployment in
resource-constrained environments, leaving them for future
work.
III. PROPOSED METHOD
A. Overview and Problem Formulation
Straw Net-ROI is an annotation -guided dual -view
classification framework for seven strawberry disease
categories: Angular Leafspot, Anthracnose Fruit Rot, Blossom
Blight, Gray Mold, Leaf Spot, Powdery Mildew Fruit, and
Powdery Mildew Leaf. The method investigate s whether
automatically selected local regions provide complementary
information to whole -image representations. It consists of a
global image classifier, a polygon -supervised ROI predictor, a
separate local -crop classifier, and confidence -adaptive
probability fusion. The proposed approach does not rely on an
intermediate assistant model or teacher –student knowledge -

Al-A'areji et al. / Journal of Applied Science and Technology Trends Vol. 07, No. 02, pp. 534-553, (2026)

538
distillation losses. Instead, the ROI and local classification
pathways are trained sequentially using the globally trained
classifier as their initialization.
For an RGB image I, the system comprises (i) a global
classifier G operating on the entire image; (ii) a region -of-
interest predictor R trained with polygon -derived masks; (iii) a
deterministic ROI crop operator C; and (iv) a separately fine -
tuned local classifier L. Each trained seed s has its own global,
ROI, and local checkpoints. All three networks are required for
the selected dual-view inference policy.
𝐳𝐺 =𝐺(𝑇(𝐼)),  𝑀=𝜎(𝑅(𝑇(𝐼))),
𝐼𝑐 =𝒞(𝐼,𝑀),  𝐳𝐿 =𝐿(𝑇(𝐼𝑐)).
(1)
In Eq. (1), T denotes bilinear resizing to 288 × 288 pixels
followed by ImageNet normalization; 𝑧𝐺 and 𝑧𝐿 are seven-class
logit vectors, M is the predicted ROI probability map, and 𝐼𝑐 is
the crop extracted from the original -resolution image. The
operator R returns ROI logits before sigmoid. Ground -truth
polygon masks supervise ROI training only; they are never used
to generate validation or test crops. Figure 1 separates the
training-only annotation pathway from inference.

Fig. 1. Architecture of Straw Net-ROI
B. Global Multiscale Classification Network
The global classification network employs an ImageNet -
1K-pretrained EfficientNet -B0 feature encoder that is fine -
tuned on the audited strawberry disease dataset. A 96 -channel
multi-scale feature pyramid and lightweight channel –spatial
attention refine its representations. Inputs are resized to 288 ×
288 pixels. Three encoder feature maps are extracted from
EfficientNet-B0 feature blocks 3, 5, and 7, comprising 40, 112,
and 320 channels, respectively. The following operations
precisely specify the residual m ulti-scale classification
network.
𝐶3=𝐸3(𝑇(𝐼)),  𝐶4=𝐸5(𝑇(𝐼)),  𝐶5 =𝐸7(𝑇(𝐼)). (2)
Ei denotes the output of the indicated EfficientNet feature
block. Lightweight channel–spatial attention (LCSA) is applied
to each Ci. Its channel gate pools spatially and applies a one -
dimensional channel convolution of kernel size three followed
by a sig moid; its spatial gate applies a 5 × 5 convolution to
concatenated channel -mean and channel -max maps. The
channel and spatial gates act sequentially.
𝑎𝑐(𝐶) =𝜎(Conv31𝐷(GAP(𝐶))),
𝑋 =𝐶⊙𝑎𝑐(𝐶),
𝑎𝑠(𝑋) =𝜎(Conv5×52𝐷 ([Avg𝑐(𝑋),Max𝑐(𝑋)])),
𝒜(𝐶) =𝑋⊙𝑎𝑠(𝑋).
(3)
In Eq. (3), σ denotes sigmoid, ⊙ element-wise
multiplication with broadcasting, GAP spatial global -average
pooling, and [·,·] channel concatenation. These established
attention operations are used as components, not claimed as
individually novel. After att ention, a 1 × 1 convolution, batch
normalization, and SiLU activation project each feature map to
96 channels. Top -down fusion uses trainable nonnegative
normalized weights rather than arbitrary unnormalized
addition.
𝑞𝑗 =
ReLU(𝑢𝑗)
∑ ReLU𝑛ℓ=1 (𝑢ℓ)+10−4,
𝒲(𝐹1,…,𝐹𝑛) =∑ 𝑞𝑗𝑛𝑗=1 𝐹𝑗. (4)
In Eq. (4), 𝑢𝑗 are trainable fusion parameters initialized to
one; 𝑞𝑗 are the nonnegative normalized weights, with a
stabilizer of 10⁻⁴. The deepest projected feature is first refined
by a depthwise -separable convolution, while the shallower
representations incorporate the bilinearly upsampled higher -
level maps through the same weighted fusion rule.
𝑃5 =𝒟(𝛱5(𝒜(𝐶5))),
𝑃4 =𝒟(𝒲(𝛱4(𝒜(𝐶4)),Up(𝑃5))),
𝑃3 =𝒟(𝒲(𝛱3(𝒜(𝐶3)),Up(𝑃4))).
(5)

Al-A'areji et al. / Journal of Applied Science and Technology Trends Vol. 07, No. 02, pp. 534-553, (2026)

539
D denotes a depth wise 3 × 3 convolution followed by a
pointwise 1 × 1 convolution, each with normalization and
SiLU, and Up denotes bilinear interpolation to the target spatial
size. Global -average pooled descriptors 𝑔𝑖 =𝐺𝐴𝑃(𝑃𝑖) each
have 96 elements. A three -output scale gate is computed using
a two -layer MLP (288 → 72 → 3) and softmax; the three
weighted descriptors are concatenated, not summed.
𝑔𝑖 =GAP(𝑃𝑖), 𝑖∈{3,4,5},
(𝑎3,𝑎4,𝑎5) =softmax(MLP([𝑔3,𝑔4,𝑔5])),
𝑔 =[𝑎3𝑔3,𝑎4𝑔4,𝑎5𝑔5].
(6)
The global logits add a baseline C5 classifier to a learned
residual from the 288 -dimensional multiscale descriptor. The
baseline is dropout (0.30) plus a 320 -to-7 linear layer. The
residual branch comprises layer normalization, dropout (0.30),
a 288-to-192 linear layer with SiLU, dropout (0.30), and a 192-
to-7 layer. A learned sigmoid scalar controls the contribution of
the residual branch.
𝐳𝐺 =𝐻5(GAP(𝐶5))+𝜎(𝜌)𝐻𝐹(𝑔). (7)
The scalar ρ is initialized to −0.7 and learned during global-
classifier training. Its fitted value is retained during inference.
Figure 2 summarizes the multi -scale feature pyramid and the
connection to the ROI prediction head.

Fig. 2. Multi-Scale Feature Extraction, Channel–Spatial Attention, Pyramid Fusion, and ROI Prediction
C. Annotation-Guided ROI Supervision
The dataset provides json polygon annotations for diseased
plant objects/ areas. Those polygons are subsequently
transformed to binary masks, 𝑀𝑖∗ . Keep in mind that these
masks should not be labeled as pixel -precise delineation of
diseased lesions (often, for example, they contain an entire
annotated leaf, fruit, or flower, not only visible lesions). An
annotation-consistency flag, 𝑣𝑖, is used to determine if a mask
qualifies for ROI supervision . The audited split retains
1,426/294/737 train/validation/test images and the same seven
image-level labels as the final comparison protocol; masks with
conflicting class annotations are excluded from the ROI loss.
For each random seed, the ROI prediction network is
initialized from the corresponding trained global classifier. The
encoder, feature pyramid, and global classification parameters
are frozen, including batch -normalization statistics; only the
ROI prediction head is optimized. The ROI head receives P3,
applies a 3 × 3 convolution (96 → 48 channels) with SiLU,
followed by a 1 × 1 convolution (48 → 1 channel). Its output is
bilinearly upsampled to 288 × 288 and passed through a
sigmoid activation to obtain the continuous ROI probability
map M.
𝑟 =Up288[Conv1×148→1(SiLU(Conv3×396→48(𝑃3)))],
𝑀 =𝜎(𝑟).
(8)
ROI training minimizes the sum of binary cross -entropy
with logits and soft Dice loss, averaged over eligible samples
𝑉=𝑖:𝑣𝑖 =1 . In Eq. (9), 𝐵𝐶𝐸𝑙𝑜𝑔𝑖𝑡𝑠(𝑟𝑖,𝑚𝑖∗) is the mean
pixelwise binary cross-entropy between raw ROI logits r_i and
binary polygon -derived target mask 𝑚𝑖∗ for image 𝑖;𝐷𝑖 is its
soft Dice coefficient. Both terms exclude samples with
inconsistent annotation labels. A batch with no valid
annotations does not update the ROI head. The Dice numerator
and denominator each use the smoothing constant 1.
ℒROI =
1
|𝒱|∑ [BCElogits(𝑟𝑖,𝑚𝑖∗)+1−𝐷𝑖]𝑖∈𝒱 ,
𝐷𝑖 =
2∑ 𝜎𝐻𝑊𝑝=1 (𝑟𝑖,𝑝)𝑚𝑖,𝑝∗ +1
∑ 𝜎𝐻𝑊𝑝=1 (𝑟𝑖,𝑝)+∑ 𝑚𝑖,𝑝∗𝐻𝑊𝑝=1 +1.
(9)
Here, 𝜎(𝑟𝑖,𝑝) denotes the probability at pixel 𝑝 and 𝑚𝑖,𝑝∗ is
the polygon -derived target. ROI -head optimization uses
AdamW (initial learning rate 2 × 10⁻⁴; weight decay 0.01),
warmup–cosine scheduling, CUDA mixed precision, a
maximum of 22 epochs, and patience seven. Validation Dice,
not classification macro -F1, determines the selected ROI
checkpoint. Spatial flips are applied consistently to images and
masks; photometric perturbations affect images alone.
D. Spatially Constrained Crop Generation
The ROI probability map is converted into a local view by
deterministic postprocessing. In Eq. (10), the 288 × 288 map is

Al-A'areji et al. / Journal of Applied Science and Technology Trends Vol. 07, No. 02, pp. 534-553, (2026)

540
smoothed using a 7 × 7 Gaussian filter with standard deviation
1.4; h is the 99.5th percentile of smoothed probabilities and b is
the 25th percentile. Maps with h < 0.52 or h − b < 0.055 receive
an explicitly logged full-image fallback because their respon se
is weak or spatially indistinct.
𝑀̃ =Gaussian7×7, 1.4(𝑀),
ℎ =𝑄0.995(𝑀̃),  𝑏=𝑄0.25(𝑀̃),
Fallback ={1, ℎ<0.52 ∨ (ℎ−𝑏)<0.055,0, otherwise.
(10)
Otherwise, component extraction evaluates q = 0.85, 0.78,
and 0.68 in that order. The binary mask uses threshold 𝜏𝑞 =
𝑚𝑎𝑥(0.50,𝑄𝑞(𝑀̃ )), where M̃ is the smoothed map. Eight -
connected components outside the allowed fractional area range
[0.002, 0.55] are discarded. For admissible components, Eq.
(11) combines positive confidence mass above b with
normalized region area. The highest -scoring component is
selected at the FIRST quantile producing admissible
candidates; scores are not compared across different quantiles.
𝑆(𝛺) =[∑ max𝑝∈𝛺 (𝑀̃𝑝−𝑏,0)]√|𝛺|𝐻𝑊,
𝛺𝑞∗ =arg max
𝛺∈𝒞𝑞
𝑆(𝛺).
(11)
The maximum-scoring component supplies the crop center
and bounding box. The box receives 14% padding per side, is
constrained to a minimum normalized side length of 0.34 and
maximum of 0.76, and is rescaled when its area would exceed
0.60 of the full image. If no admissible component is found but
the probability map has adequate contrast, the algorithm
searches 46%- and 60%-wide confidence windows and uses the
strongest window; otherwise , the full -image fallback is
recorded. Coordinates are mapped back to the original image
dimensions, cropped with PIL, and resized to 288 × 288 pixels.
The quality-control gate requires at least 55% of training and
validation crops to be localized (area fraction < 0.82); it blocks
model training if this requirement fails. Ground-truth masks are
never consulted to choose a crop box.
This step highlights a high -response spatially contiguous
crop, while discouraging overly expansive ones. It is a region -
selection heuristic based on the predicted ROI maps, but the
provided annotation and predicted regions are not meant to
precisely annot ate lesion extent to the pixel level. Crop
metadata and qualitative previews are retained for visual
inspection.
E. Global and Local Classifier Training
Global-classifier training. For each of five random seeds
(42, 43, 44, 45, and 46), the global network is optimized using
288 × 288 images, batch size 16, up to 28 epochs, and early -
stopping patience 10. The optimizer is AdamW, with initial
learning rates of 7 × 10⁻⁶ for the encoder and 4 × 10⁻⁵ for task -
specific layers. An exponential moving average of model
weights (decay 0.995) is used for checkpoint selection. Once
trained, the validation -selected global classifier is retained
unchanged during optimization of the spatial and local modules.
ROI-predictor training. The predictor is initialized from the
corresponding global classifier, whose parameters remain
frozen while the ROI head is trained on polygon-derived masks
as described in Section III.C. The best ROI predictor is selected
using val idation Dice. Its predictions generate training and
validation crops, and the same deterministic procedure
generates test -time crops from images alone. Ground -truth
polygons are never used to provide crop coordinates during
inference.
Local-classifier fine -tuning. A separate EfficientNet -
B0/FPN/LCSA classifier is initialized using the corresponding
trained global weights. Encoder blocks 0–5 are frozen and kept
in evaluation mode. Later encoder blocks, attention, projection,
pyramid-fusion, and classification layers are optimized using
automatically generated ROI crops. The local training objective
is class-weighted cross-entropy with 0.025 label smoothing. For
training count 𝑛𝑐 in class c and mean class count 𝑛̄ , weights are
proportional to (𝑛̄ /𝑛𝑐)0.25 and normalized to have a mean of
one across classes.
𝑛‾ =
1
𝐾∑ 𝑛𝑗𝐾𝑗=1 ,  𝐾=7,
𝑤𝑐 =
(𝑛‾/𝑛𝑐)0.25
1
𝐾∑ (𝑛‾/𝑛𝑗)0.25𝐾𝑗=1 ,
𝑡𝑖,𝑐 =(1−𝜀) 𝟏[𝑦𝑖 =𝑐]+ 𝜀𝐾, 𝜀=0.025,
ℒlocal =−
∑ ∑ 𝑤𝑐𝐾𝑐=1𝐵𝑖=1  𝑡𝑖,𝑐ln𝑝𝐿,𝑖,𝑐
∑ 𝑤𝑦𝑖𝐵𝑖=1 .
(12)
Equation (12) defines 𝑛𝑐 as the training count for class c, n̄
as the mean class count, 𝑤𝑐 as the normalized inverse -
frequency weight, and 𝜀=0.025 as the label -smoothing
coefficient. The target 𝑡𝑖,𝑐 assigns probability (1−ε) to the true
label plus 𝜀/𝐾 uniformly across classes; 𝑝𝐿,𝑖,𝑐 =
𝑠𝑜𝑓𝑡𝑚𝑎𝑥(𝑧𝐿,𝑖)𝑐. For a mini -batch of B images, the weighted
loss is normalized by the sum of the true -class weights,
matching the weighted -mean cross -entropy implementation.
AdamW uses weight decay 0.015 and initial learning rates 1.2
× 10⁻⁵ for late encoder blocks, 3.0 × 10⁻⁵ for FPN/attention, and
8.0 × 10⁻⁵ for heads and gates. Batch size is 20, maximum
epochs 25, and patience is seven; the highest validation macro -
F1 checkpoint is retained. Training uses warmup –cosine
scheduling, seeded augmentations, gradient clippin g, mixed
precision, and resumable optimizer/RNG state.
F. Confidence-Adaptive Decision Fusion
The two trained classifiers produce separate probability
vectors 𝑝𝐺 =𝑠𝑜𝑓𝑡𝑚𝑎𝑥(𝑧𝐺) and 𝑝𝐿 =𝑠𝑜𝑓𝑡𝑚𝑎𝑥(𝑧𝐿) . Their
maximum class probabilities 𝑐𝐺 and 𝑐𝐿 are used as heuristic
confidence indicators, not calibrated uncertainty estimates. The
selected confidence -adaptive fusion coefficient 𝑤 varies with
the difference between local and global confidence:

Al-A'areji et al. / Journal of Applied Science and Technology Trends Vol. 07, No. 02, pp. 534-553, (2026)

541
𝑐𝐺 = max1≤𝑘≤𝐾𝑝𝐺,𝑘,  𝑐𝐿 = max1≤𝑘≤𝐾𝑝𝐿,𝑘,
𝑤 =𝜎[ln(
𝛼0
1−𝛼0
)+𝛽(𝑐𝐿−𝑐𝐺)]. (13)
The validation -selected fusion configuration uses 𝛼₀=
0.30 and β = 5.0. The quantity 𝜎(𝑎)=1/(1+𝑒𝑥𝑝(−𝑎)) is
the logistic sigmoid, and w is the sample -specific weight
assigned to local predictions. Therefore, 𝑐𝐺 =𝑐𝐿 implies 𝑤=
𝛼₀ . The local contribution rises when 𝑐𝐿 exceeds 𝑐𝐺 . These
softmax maxima are heuristic confidence values rather than
calibrated uncertainty estimates; fusion is applied to
probabilities rather than feature maps.
𝑝final,𝑘 =(1−𝑤)𝑝𝐺,𝑘+𝑤𝑝𝐿,𝑘, 𝑘=1,…,𝐾,
𝑦̂ =arg max
1≤𝑘≤𝐾
𝑝final,𝑘,  𝐾=7. (14)
The candidate set consisted of global-only, crop-only, fixed
local weights {0.25, 0.50, 0.75}, and confidence -dependent
fusion with 𝛼0∈0.30,0.50,0.70 and 𝛽∈2,5,10. One common
policy was selected by the highest mean validation macro -F1
across all five seeds, with ties favoring the earlier, simpler
policy. Thus the selected policy is not refitted separately for
each test sample or seed. No GT masks or test labels ent er the
fusion equation.
G. Experimental Protocol, Ablations, and Reproducibility
The source archive comprises 2,500 images across seven
classes. Following exact -duplicate removal and exclusion of
two ambiguous records, the audited partition includes 2,457
images (1,426 training, 294 validation , and 737 testing). The
deterministic partition identifier is 4f7ee279d8bd0690, and the
original archive SHA -256 fingerprint is
e728dd56eaab4e11e64d46da77b06eb377beb264c53d0a32c1ea
f44d051de525. Exact -duplicate and perceptual -similarity
screening preceded m odel comparison; however, plant -level
and acquisition-session independence cannot be independently
established.
Classification performance is characterized by accuracy,
macro-precision, macro -recall, and macro -F1 over all seven
classes. In Eq. (15), 𝑃𝑘 =𝑇𝑃𝑘/(𝑇𝑃𝑘+𝐹𝑃𝑘) and 𝑅𝑘 =𝑇𝑃𝑘/
(𝑇𝑃𝑘+𝐹𝑁𝑘); if a precision, recall, or F1 denominator is zero,
its value is defined as zero, matching the evaluation code. ROI
Dice/IoU are calculated only for eligible polygon -derived
masks and do not measure pixel -accurate lesion boundaries.
Ablations include glob al-only, ROI -crop-only, fixed and
confidence-adaptive fusion, plus center - and random -crop
alternatives. For the latter, crop sizes are matched to ROI -
generated boxes, so these are spatial controls rather than
completely independent crop-size strategies.
Macro−F1=
1
𝐾∑ 2𝑃𝑘𝑅𝑘𝑃𝑘+𝑅𝑘𝐾𝑘=1 ,  𝐾=7. (15)
The complete inference system consists of 4,165,681
parameters in the global classifier, 4,245,722 in the ROI
predictor, and 4,165,681 in the local classifier, totaling
12,577,084 parameters. In a batch -one end-to-end evaluation,
confidence-adaptive fusion required 18.80 ms per image
compared with 6.70 ms for global -only inference on an
NVIDIA RTX PRO 6000 Blackwell Server Edition (288 × 288
input, 60 images, one seed). These metrics include loading,
preprocessing and inference on images, but does not include
loading the model -checkpoint. Furthermore, there is no
indication of the embedded -device latency and energy
efficiency. The memory allocated from the PyTorch peak
memory is probably not equivalent to the total occupation of the
GPU.
The 737-image test partition was sampled multiple times
during the testing process of the methodology. While the
reported fusion configuration was selected based on validation
evidence, the test estimates reflect exploratory internal testing
rather than in dependent evidence of generalization. The
proposed comprehensive framework outperformed using only
global features across five random seeds with an average test
macro-F1 of 97.79% (Standard deviation 0.47 percentage
points) versus 97.20% for the classifier based on global features
alone. The observed increase was not statistically significant (p
= 0.127) and other center- and random-crop validation controls
did not demonstrate statistically significant superiority of ROI -
fusion after correction for multiple comparisons. These
limitations advise caution when interpreting claims concerning
the spatial branch's contribution.
IV. EXPERIMENTAL SETUP
This section describes the data set, preprocessing and
quality-control approaches, measures of performance,
implementation parameters and experimental design used in the
assessment of Straw Net -ROI. All the experiments described
here are based on the fixed validated data set split, and the results
are presented with regard to the limitations of the internal
evaluation protocol.
A. Dataset Description
The Dataset origin and acquisition conditions. The
experiments were conducted using the publicly available
Strawberry Disease Detection Dataset introduced by Afzaal et
al. [19]. The original dataset contains 2,500 RGB images
covering seven strawberry disease categories: Angular Leafspot,
Anthracnose Fruit Rot, Blossom Blight, Gray Mold, Leaf Spot,
Powdery Mildew Fruit, and Powdery Mildew Leaf. The
published dataset was originally divided into 1,450 training, 307
validation, and 743 testing images.
According to Afzaal et al. [4], the majority of the images
were acquired from multiple greenhouses in South Korea using
camera-equipped mobile phones under natural illumination,
while approximately 20% were collected from publicly available
agricultural and plant pathology sources. The dataset contains
variations in imaging distance, illumination, background
complexity, and disease development stages. The released
images have a spatial resolution of 419 × 419 pixels,
representing their processed dataset resolution rather than
necessarily the native camera resolution. No additional field
images were collected for the present study. Plant identities,

Al-A'areji et al. / Journal of Applied Science and Technology Trends Vol. 07, No. 02, pp. 534-553, (2026)

542
individual acquisition-session identifiers, and complete image -
level source provenance were unavailable in the dataset copy
used for the experiments.

  1. Annotation procedure and quality control
    The original dataset authors annotated disease -associated
    regions using the Labelme annotation tool [19]. The distributed
    dataset includes image -associated JSON files containing
    polygon vertices and disease labels. Our initial inspection
    identified 5,707 p olygon annotation records across the 2,500
    source images. The polygons were converted into binary masks
    for supervised ROI prediction. Importantly, the annotations were
    considered regions with disease inside a plant object rather than
    exact pixel -based lesion outlines – some polygons cover an
    entire infected leaf, flower or fruit rather than only the visible
    symptom.
    Annotation consistency was assessed by testing the labels on
    the image against labels on the polygon. Six of the remaining
    images had inconsistencies: four in the training dataset, one in
    the validation dataset, and one in the testing dataset. The
    inconsistent images were not used for spatial supervision if there
    was a mismatch between polygon annotations, although their
    image-level label was saved to be used for classification. There
    was an issue with annotation consistency in our dataset, which
    may require confirmation from the original creators.
  2. Duplicate screening and group-aware partitioning
    The original dataset partition was subjected to a
    deterministic quality -control procedure before final model
    evaluation. Exact duplicate records were identified using SHA-
    256 image hashes. Potentially related images were additionally
    screened using 64 -bit difference hashing (dHash ), with a
    Hamming-distance threshold of four. Image pairs satisfying the
    similarity criterion were joined into connected components,
    producing groups of potentially related images. These groups
    were treated as indivisible units during dataset partitioning.
    A class -stratified, group -aware allocation procedure was
    applied with a fixed partition seed of 42 and target proportions
    of 58%, 12%, and 30% for training, validation, and testing,
    respectively. Whole similarity groups, rather than individual
    images, were assigned to partitions. Groups containing
    conflicting class labels were quarantined, and exact duplicate
    copies were removed through deterministic canonical -record
    selection.
    The quality-control procedure excluded 41 exact-duplicate
    records and quarantined two records belonging to conflicting -
    label groups. The resulting audited dataset contained 2,457
    images, divided into 1,426 training, 294 validation, and 737
    testing images.
    The retained dataset comprised 2,154 distinct similarity
    groups. A check of the retained manifest indicates that the total
    of 2,457 images all has unique SHA-256 signatures, and that no
    similarity group observed exists in multiple partitions. This
    evidence shows that the groups are disjoint at the level of image
    similarity defined by the used procedure. But it does not ensure
    biological plant-level nor acquisition-session independence.
  3. Class-wise distribution
    Table II presents the distribution of the audited dataset across
    all seven disease categories. The partition was fixed for
    subsequent experiments and model comparisons.
    TABLE II. CLASS-WISE DISTRIBUTION OF THE AUDITED
    STRAWBERRY DISEASE DATASET
    Disease Class Train Validation Test Total
    Angular Leafspot 252 52 131 435
    Anthracnose Fruit Rot 56 12 29 97
    Blossom Blight 121 25 62 208
    Gray Mold 276 57 143 476
    Leaf Spot 334 68 172 574
    Powdery Mildew Fruit 78 16 40 134
    Powdery Mildew Leaf 309 64 160 533
    Total 1,426 294 737 2,457
    The resulting class distribution remained imbalanced. Leaf
    Spot and Powdery Mildew Leaf represented relatively frequent
    categories, whereas Anthracnose Fruit Rot and Powdery Mildew
    Fruit were substantially less represented. Macro -averaged
    precision, recall , and F1 were therefore included to evaluate
    classification performance without allowing the largest classes
    to dominate the principal reported metric.
  4. Reproducibility and generalisation limitations
    The audited dataset was identified by the fixed split
    fingerprint 4f7ee279d8bd0690, while the original source archive
    was identified by SHA -256 checksum
    e728dd56eaab4e11e64d46da77b06eb377beb264c53d0a32c1ea
    f44d051de525. The source archive was preserved unch anged,
    and the reconstructed partition was used consistently in the final
    experimental protocol.
    The similarity -based grouping process implies a lower
    chance of exact duplicate and detectable near-duplicate leakage.
    However, dHash similarity seems unlikely to act as a direct
    substitute to a validated identifier, such as plant, greenhouse,
    source or acquisition-session identifier. As a result, snapshots of
    the same plant or acquisition session cannot be safely assumed
    to be adequately separated into different subsets.
    Furthermore, the group-aware partition was assembled back
    in time from available source images, and test source images
    from history were evaluated during early stages. The final test
    partition of 737 images is therefore not guaranteed to be fully
    non-overlapping or prospective. External evaluation was also
    proposed with a second public strawberry dataset, but taxonomy
    of classes within this external dataset was not compatible with
    either seven-class classification task to perform a final seven -
    class evaluat ion. Purchase provenance verification and cross -
    dataset near -duplicate resolution were also unavailable .
    Therefore, no independent external seven -class classification
    performance is reported.

Al-A'areji et al. / Journal of Applied Science and Technology Trends Vol. 07, No. 02, pp. 534-553, (2026)

543
The evaluation results are consequently interpreted as
exploratory internal estimates of classification performance.
They do not establish generalisation to independently collected
greenhouses, unseen acquisition sessions, different geographic
regions, or operational field deployments. Such claims require
further evaluation on independently collected and appropriately
annotated external data.
B. Evaluation Metrics
The evaluation of model performance requires multiple
metrics, which focus on macro -F1 score as the primary metric
for assessment of imbalanced classification tasks.
The accuracy metric is defined as:
Accuracy=
𝑇𝑃+𝑇𝑁
𝑇𝑃+𝑇𝑁+𝐹𝑃+𝐹𝑁 (53)
The overall correctness of a system can be evaluated
through accuracy, yet this metric shows bias toward the
dominant classes when used with imbalanced datasets. The
macro-F1 score serves as the main evaluation metric because it
enables to solve the proble m that exists in this method. The
system calculates F1 scores for all classes and then determines
the unweighted average of those scores in the following way:
Macro-F1=
1
𝐾∑
2⋅𝑃𝑘⋅𝑅𝑘
𝑃𝑘+𝑅𝑘
𝐾
𝑘=1
(54)
Where 𝑃𝑘and 𝑅𝑘denote precision and recall for the class 𝑘.
The metric treats all classes equally and appropriately elucidates
performance on minority classes.
In addition, macro-precision and macro-recall are reported:
Precision=
𝑇𝑃
𝑇𝑃+𝐹𝑃,
Recall=.
𝑇𝑃
𝑇𝑃+𝐹𝑁

(55)
The metrics present additional evidence which
demonstrates how well the model performs at disease class
identification while maintaining its accuracy in avoiding
incorrect predictions.
C. Implementation Details

  1. Experimental Environment and Reproducibility
    The proposed Straw Net-ROI framework was implemented
    using Python and Py Torch with the torchvision EfficientNet-B0
    architecture. The experiments were executed in a GPU-enabled
    Google Colab environment. Five fixed random seeds (42, 43, 44,
    45, and 46) were used to evaluate training variability. Python,
    NumPy, and Py Torch random -number generators were
    initialized using the corresponding seed. CUDA random -
    number generators were also seeded, while cuDNN
    benchmarking was disabled and deterministic cuDNN
    operations were requested to reduce nondeterministic behavior.
    All experiments used the same audited dataset partition,
    comprising 1,426 training images, 294 validation images, and
    737 testing images. The dataset partition was identified by the
    split fingerprint 4f7ee279d8bd0690. The original dataset archive
    was verified using its recorded SHA-256 checksum.
    RGB images were normalized using ImageNet channel
    statistics, with mean values (0.485, 0.456, 0.406) and standard
    deviations (0.229, 0.224, 0.225). Mixed -precision training was
    enabled on CUDA -compatible hardware. Optimizer states,
    model parameters, gradie nt-scaler states, and random -number
    generator states were retained in resumable checkpoints.
  2. Global Classification Network
    The global classification network uses the torchvision
    EfficientNet-B0 feature extractor initialized with the
    EfficientNet_B0_Weights.IMAGENET1K_V1 pretrained
    weights. Three intermediate feature representations are extracted
    from encoder feature blocks 3, 5, and 7, producing feature maps
    with 40, 112, and 320 channels, respectively.
    Each feature map is processed by a lightweight channel –
    spatial attention (LCSA) module. Channel attention is
    implemented through global average pooling followed by a one-
    dimensional convolution with kernel size three and sigmoid
    activation. The spatial attention stage concatenates channel-wise
    average and maximum feature responses and applies a 5 × 5
    convolution followed by sigmoid activation.
    The attention-refined features are projected into a common
    96-channel representation using 1 × 1 convolution, batch
    normalization, and SiLU activation. The feature pyramid uses
    top-down bilinear interpolation and learned nonnegative fusion
    weights, normalized with a numerical stabilization constant of
    10⁻⁴. Depthwise -separable convolution refines the resulting
    pyramid representations.
    Global average pooling produces three 96 -dimensional
    feature vectors. A scale -aware gating network combines these
    representations using a two -layer multilayer perceptron with a
    hidden dimension of 72 and softmax-normalized scale weights.
    The resulting 288 -dimensional descriptor is processed by the
    fusion classification head. A residual contribution from the
    deepest encoder feature map is retained to preserve the original
    global classification pathway.
    A dropout rate of 0.30 is used in the classification heads.
    The fusion residual is controlled by a learnable sigmoid gate
    initialized with a logit value of −0.7.
  3. Global Classifier Training
    Training of the global classifier is performed in two
    sequential phases.
    First, the ImageNet -pretrained EfficientNet -B0-based
    classification architecture is trained using 224 × 224 images. The
    maximum training duration is 100 epochs, with batch size 32 and
    early-stopping patience 25. AdamW is employed with initial
    learning rates of 4 × 10⁻⁵ for encoder parameters and 4 × 10⁻⁴ for
    task-specific parameters, and weight decay of 10⁻⁴. The
    learning-rate schedule includes five warmup epochs followed by
    cosine decay to a minimum learning rate of 10⁻⁶.
    An exponential moving average (EMA) of model
    parameters is maintained using a decay coefficient of 0.999.
    Validation checkpoint selection uses the original model during
    the initial three epochs and the EMA model thereafter. The
    checkpoint with the highest validation macro-F1 is retained.

Al-A'areji et al. / Journal of Applied Science and Technology Trends Vol. 07, No. 02, pp. 534-553, (2026)

544
The training objective is class-weighted focal cross-entropy
with focal parameter γ = 1.5 and label smoothing ε = 0.05. Class
weights are calculated from training -set frequencies using an
inverse-frequency exponent of 0.5.
Second, the selected global model is further fine -tuned
using 288 × 288 inputs for a maximum of 28 epochs, batch size
16, and patience 10. The encoder and task -specific layers use
initial learning rates of 7 × 10⁻⁶ and 4 × 10⁻⁵, respectively.
AdamW weight decay remains 10⁻⁴. A two-epoch warmup and
cosine schedule are applied, with a minimum learning rate of
10⁻⁷.
During this phase, EMA uses a decay coefficient of 0.995.
Validation macro -F1 determines the selected checkpoint, and
gradient clipping uses a maximum norm of 2.0.
After global -classifier training, the validation -selected
checkpoint for each seed is retained as the reference initialization
for subsequent ROI and local-classification training.
4. ROI Prediction Training
The training of the ROI prediction network is initialized
with the corresponding trained global classification weights.
During training of the ROI supervision, the feature encoder,
multi-scale pyramid, and classification parameters are fixed.
Additionally, batch-normalization statistics within these fixed
blocks are copied from training mode to evaluation mode. The
96-channel P3 feature map is passed through the ROI head as a
3×3 convolution that produces 48 channels, which go through
SiLU activation, and then through a 1×1 convolution, which is
presumably a single -channel ROI logit map . Bilinear
interpolation restores the prediction to 288 × 288 pixels.
Supervision for learning ROI is performed with polygon -
derived binary masks. The loss combines both binary cross -
entropy with logits, soft Dice loss with equal coefficients. Only
samples with suitable spatial annotations can possibly impact the
ROI objective. The Dice smoothing constant is kept to 1. The
ROI head is trained with AdamW optimizers with 2 10,4
learning rate and 0.01 weight decay. With batch size 16,
maximum 22 epochs, warmup -cosine schedule, early -stopping
patience 7, and gradient clipping 1.0 (maximum norm), the
checkpoint with the highest validation Dice score is used f or
training. Horizontal and vertical flips are used consistently in
training image and its mask annotation. Photometric data
augmentation is applied solely on input image and not on mask
annotations.
5. ROI Crop Generation
At inference time, the trained ROI predictor is used to
predict an ROI probability map from the input image. This map
is smoothed with a Gaussian kernel (size 7 × 7, σ 1.4). The
studies show that the 99.5th and 25th percentile of the smoothed
map are used to determine the ROI confidence and spatial
contrast, respectively. A full image fallback is probably used
when the upper -confidence statistic is less than 0.52 or the
confidence contrast is less than 0.055.
In the case of eligible maps, connected components are
examined with map quantile thresholds of 0.85, 0.78, and 0.68
tested in order. The corresponding pixel threshold may be
selected as the maximum of 0.50 and the selected map quantile.
Eight-connected components with normalized area outside
[0.002, 0.55] are disregarded. Eligible components are ordered
according to the positive confidence mass and normalized
spatial extent.
The selected component determines the crop center and
bounding box. A padding fraction of 0.14 is applied on each side,
and normalized crop dimensions are constrained to the interval
[0.34, 0.76]. Crop areas greater than 0.60 of the complete image
are redu ced. If no admissible component is found but the
confidence map remains spatially informative, confidence -
weighted windows covering approximately 46% or 60% of the
image dimensions are examined. Otherwise, the complete image
is used as a fallback. The extracted crop is resized to 288 × 288
pixels. The procedure relies exclusively on predicted ROI maps
for determining crop coordinates; ground -truth polygon masks
are not used for validation or testing crop selection.
6. Local Classification Network
The local classification network has the same EfficientNet-
B0, LCSA, and multi -scale fusion architecture as the global
classifier. It is initialized using the corresponding global
classification checkpoint. Encoder feature blocks 0–5 are frozen,
while later encoder blocks, attention modules, feature pyramid
layers, and classification heads are fine -tuned using predicted
ROI crops.
Training uses AdamW with three parameter groups: 1.2 ×
10⁻⁵ for the later encoder blocks, 3 × 10⁻⁵ for attention and
feature pyramid components, and 8 × 10⁻⁵ for classification
heads and gating parameters. Weight decay is set to 0.015. The
local classifier is trained for up to 25 epochs with batch size 20
and early -stopping patience seven. A two -epoch warmup and
cosine-decay schedule is applied, using the learning-rate scaling
procedure implemented in the training code. Gradient clipping
uses a maximum norm of 1.0.
The classification objective is class-weighted cross-entropy
with label smoothing ε = 0.025. Class weights are derived from
the training -set class frequencies using an inverse -frequency
exponent of 0.25 and normalized to have a mean of one across
classes. Model selection is based on validation macro -F1. The
best local-classifier checkpoint is retained for each random seed.
7. Data Augmentation
For initial and subsequent global-classifier training, images
are augmented using random resized cropping, horizontal
flipping, vertical flipping, Rand Augment, and random erasing.
Random resized cropping uses a scale interval of [0.70, 1.00]
and aspect-ratio interval of [0.85, 1.15]. Horizontal flipping has
probability 0.50, vertical flipping has probability 0.15, and
random erasing has probability 0.20. Rand Augment is
configured with two operations and magnitude nine.
For ROI-predictor training, horizontal and vertical flipping
probabilities are 0.50 and 0.30, respectively. Color adjustment is
applied with probability 0.50 and multiplicative range [0.85,
1.15], while brightness adjustment is applied with probability
0.50 and range [0.88, 1.12]. For local -classifier training,
horizontal and vertical flipping probabilities are 0.50 and 0.20,
respectively. Color adjustment is applied with probability 0.50
and multiplicative range [0.88, 1.12], while brightness

Al-A'areji et al. / Journal of Applied Science and Technology Trends Vol. 07, No. 02, pp. 534-553, (2026)

545
adjustment is applied with probability 0.50 and range [0.90,
1.10]. No stochastic data augmentation is applied during
validation or testing.
8. Confidence-Adaptive Prediction Fusion
The global and local classification networks produce
separate seven-class soft max probability distributions. Their
maximum predicted class probabilities are used as confidence
indicators for sample -dependent fusion. Candidate
configurations include global -only classification, local -only
classification, fixed local-branch contributions of 0.25, 0.50, and
0.75, and confidence -adaptive fusion. Confidence-adaptive
configurations are evaluated using nominal local-branch weights
of 0.30, 0.50, and 0.70 and confidence-sensitivity parameters of
2, 5, and 10.
A single fusion configuration is selected by maximizing the
mean validation macro -F1 across all five random seeds. The
selected configuration uses a nominal local -branch weight of
0.30 and confidence sensitivity of 5.0. This configuration is
applied consistently to the final reported classification runs.
9. Distillation and Model Compression
No teacher –assistant–student knowledge -distillation
procedure is used in the proposed StrawNet -ROI framework.
Accordingly, there is no intermediate assistant model,
distillation temperature schedule, feature -distillation loss, or
relational-distillation objective associated with the final method.
Likewise, quantization and pruning are not part of the evaluated
implementation.
The EMA procedure described above is used exclusively as
a training-stabilization and checkpoint-selection mechanism for
the global classification network, not as a form of knowledge
distillation. The reported method therefore consists of global -
classifier training, ROI-head supervision, local -classifier fine-
tuning, and validation -selected confidence-adaptive prediction
fusion.
D. Results
The proposed StrawNet-ROI framework was evaluated on
the audited seven -class strawberry disease dataset, comprising
1,426 training images, 294 validation images, and 737 testing
images. Five experimental runs were conducted using fixed
random seeds (42–46). Classification performance was assessed
using accuracy, macro -precision, macro-recall, and macro -F1,
with macro -F1 serving as the primary metric because of the
unequal class distribution.

  1. Overall Classification Performance
    Across five independent training runs using fixed random
    seeds (42–46), Straw Net-ROI achieved a mean test accuracy of
    98.78 ± 0.25%, macro-precision of 98.26 ± 0.60%, macro-recall
    of 97.40 ± 0.43%, and macro -F1 of 97.79 ± 0.47%, where the
    uncertainty values represent sample standard deviations across
    runs.
    Table III presents per-run values for each of the four main
    metrics, as well as their means and sample standard deviations.
    The single best macro -F1 value across the total 4×5 runs was
    achieved by seed 45 (macro -F1 = 98.41%) (test accuracy =
    99.19%). This best result for the individual run is presented here
    apart from the 5 -run average and is not used as the overall
    performance estimate.
    Using the same audited test partition and random seeds, the
    global-only classifier's mean test accuracy and macro-F1 across
    the 10 predetermined random seeds were 98.56% and 97.20 ±
    0.60%, respectively. Based on these results, it is anticipated that
    the du al-view setting may have realized incremental gains of
    approximately 0.22 percentage points in mean accuracy, and
    0.59 percentage points in mean macro-F1. The paired macro-F1
    figures for the two configurations are also depicted in Fig. 3.
    Because the same 737 -image test partition was evaluated
    across all five seeds, the reported standard deviations
    characterize variation associated with model training rather than
    independent test -sample uncertainty. Moreover, the historical
    test partition h ad been examined during methodological
    development; consequently, these results are interpreted as
    exploratory internal performance estimates.
    TABLE III. PER-RUN AND AGGREGATED TEST CLASSIFICATION PERFORMANCE OF STRAWNET-ROI
    Seed Accuracy (%) Macro-Precision (%) Macro-Recall (%) Macro-F1 (%)
    42 98.6431 98.0773 96.8077 97.3686
    43 98.7788 98.5981 97.5544 98.0555
    44 98.7788 98.3057 97.4189 97.8414
    45 99.1859 98.9405 97.9758 98.4150
    46 98.5075 97.3579 97.2191 97.2809
    Mean 98.7788 98.2559 97.3952 97.7923
    Sample SD 0.2538 0.5972 0.4299 0.4745
    The proposed framework achieved higher macro-F1 in four
    of the five runs, although one run showed a decrease relative to
    the global -only classifier. Therefore, the observed benefit of
    integrating global and localized representations was not uniform
    across all random seeds.
  2. Statistical Assessment
    A Statistical variability was assessed across five
    independent training runs conducted with fixed random seeds of
    42, 43, 44, 45, and 46. For each principal classification metric,

Al-A'areji et al. / Journal of Applied Science and Technology Trends Vol. 07, No. 02, pp. 534-553, (2026)

546
the arithmetic mean and sample standard deviation were
calculated from the corresponding seed -level test results.
Exploratory two-sided 95% confidence intervals for the mean
were estimated using the Student's t -distribution with four
degrees of freedom.

Fig. 3. Test Macro-F1 Scores of Straw net-ROI and the Global-Only
Classifier Across Five Seeds
StrawNet-ROI achieved a mean test accuracy of 98.78 ±
0.25%, macro-precision of 98.26 ± 0.60%, macro-recall of 97.40
± 0.43%, and macro-F1 of 97.79 ± 0.47%, where the values after
the ± symbol represent sample standard deviations in percentage
points. The explorat ory 95% confidence interval for the mean
macro-F1 was 97.20–98.38%.
A paired statistical analysis was performed against the
global-only classifier, which provides the most directly
controlled comparison for evaluating the additional ROI -based
pathway. Both configurations were evaluated using matched
random seeds and the sa me 737 test images. The mean paired
macro-F1 difference was 0.5929 percentage points, with a
sample standard deviation of 0.6893 percentage points and an
exploratory 95% confidence interval of −0.2630 to +1.4488
percentage points.
The two-sided paired t -test did not indicate a statistically
significant difference ( 𝑡(4)=1.9233,𝑝=0.1268 ). A
complementary Wilcoxon signed -rank test likewise yielded a
nonsignificant result (p = 0.1875). The proposed framework
achieved higher macro-F1 scores in four of the five paired runs,
but its numerical advantage was not consistent across all seeds.
Means, sample standard deviations and exploratory
confidence intervals for the other comparative architectures are
shown in a descriptive fashion. Formal paired significance
claims are not made for these comparisons because no complete
seed-level matched records exist for independent verification of
them in this study. Additionally, differences in training
configurations limit a controlled attribution of performance
differences.
The statistical conclusions are subject to several limitations.
First, five training seeds provide only a limited estimate of
training variability and yield relatively wide confidence
intervals. Second, the same test images were evaluated
repeatedly across seeds; therefore, the estimated confidence
intervals do not characterize uncertainty across independently
sampled test populations. Third, the historical test partition was
inspected repeatedly during methodological development.
Accordingly, all test-based statistical findings are interpreted as
exploratory internal evidence rather than independent
confirmatory demonstrations of generalization or architectural
superiority.
3. Class-Wise Classification Performance
To examine whether the overall classification performance
extends to underrepresented disease categories, we report class-
wise precision, recall, and F1-score for all seven categories. The
analysis was independently recomputed from the saved
predictions of the final StrawNet-ROI framework for five fixed
random seeds (42–46). The same audited test partition of 737
unique images was used for each seed, and class-wise measures
were calculated separately for each run before averaging. Table
IV therefore reports the mean performance across five runs, with
standard deviations for class-wise F1 and the number of distinct
test images per class.
TABLE IV. CLASS-WISE TEST PERFORMANCE OF STRAWNET-ROI ACROSS FIVE RANDOM SEEDS
Disease class Precision (%) Recall (%) F1-score (%) Support
Angular Leafspot 99.55 100.00 99.77 ± 0.21 131
Anthracnose Fruit Rot 97.03 88.97 92.79 ± 2.91 29
Blossom Blight 99.68 100.00 99.84 ± 0.36 62
Gray Mold 97.64 98.04 97.84 ± 0.68 143
Leaf Spot 98.85 99.88 99.36 ± 0.13 172
Powdery Mildew Fruit 95.05 95.50 95.26 ± 1.38 40
Powdery Mildew Leaf 100.00 99.38 99.69 ± 0.00 160
Macro average 98.26 97.40 97.79 ± 0.47 737
The results demonstrate heterogeneous performance across
disease categories. Angular Leafspot, Blossom Blight, Leaf Spot
and Powdery Mildew Leaf, all of which had a mean F1-score of
over 99%. The lowest mean F1 -score and recall were obtained

Al-A'areji et al. / Journal of Applied Science and Technology Trends Vol. 07, No. 02, pp. 534-553, (2026)

547
by the category Anthracnose Fruit Rot, the smallest data
category with 29 samples, with an F1 -score of 92.79% and a
recall of 88.97%. This indicates that the category "Anthracnose
Fruit Rot" may be the most difficult to classify correctly.
Powdery Mildew Fruit produced an F1 -score of 95.26% and a
recall of 95.50%, with 40 samples in the category. It is not
necessary to draw conclusions about the minority -class results,
as very small changes will result in the class -wise estimates
changing significantly.
Based on the class -wise results, it can be inferred that
achieving high overall classification accuracy may not mean
consistency for all the seven disease groups. It can be seen that
the difference between the mean macro-precision (98.26%) and
mean macro -recall (97.40%) metrics indicates their non -
equivalence. The largest precision–recall disparity was observed
for Anthracnose Fruit Rot, with precision of 97.03% and recall
of 88.97%.
Numerical results show that using the class-specific internal
test set instead of the global -only classifier Straw Net -ROI
improved macro-F1 by 1.55 percentage points on Anthracnose
Fruit Rot, and 2.63 percentage points on Powdery Mildew Fruit.
These class-specific differences may be affected by the sampling
variability, because they are obtained over the same historically
re-used internal test partition.
The differences we observed cannot be explained solely by
the class-weighted focal objective because the global -only and
dual-view configurations share the same global classification
branch, which is optimised using the focal loss. In order to
disentangle the effect of focal reweighting, one would need to
compare the outcomes across two models that differ only in their
loss formulation.
In addition, Figure 4 presents the combined confusion
matrix over 5 individual seeds. For each true class, the cell
percentages are normalized by the total number of predictions
(aggregated over all five runs). Note that the individual data
points are 3,685 total predictions (737 images x five repetitions)
instead of 3,685 distinct test images, so the matrix here shows
repeated prediction behavior, not independent sample -based
uncertainty estimates.

Fig. 4. Row-Normalized Test Confusion Matrix of Strawnet-ROI Aggregated Over Five Seeds
The most frequent errors were the ones between fruit
diseases that look similar. Powdery Mildew Fruit was identified
as Gray Mold in 9 of the predictions for 5 of the seed sets and
vice versa; Gray Mold was identified as Powdery Mildew Fruit
in 7 of the pr edictions for 5 of the seed sets. Anthracnose Fruit
Rot was identified as Gray Mold 7 times and Leaf Spot 5 times.
Angular Leafspot and Blossom Blight were both perfect. The
errors falling in small classes and visually similar diseases may
indicate that it is probably not entirely fair to say that disease
recognition on average is always reliable.
These results are from the internal test on a historical
partition that has been re -tested in the process of method
development. Results should not be taken as independent
evidence of field -level generalization, nor as evidence that the
proposed system reliably makes successful decision on difficult
minority-class cases in new acquisition conditions.
4. Computational Characteristics
To quantify the inference efficiency of the Straw Net-ROI
pipeline, a batch-one benchmark was performed on an NVIDIA
RTX PRO 6000 Blackwell Server Edition GPU with PyTorch.
The benchmark consisted of performing inference on RGB
images, scaled to 288 × 288 pixels. We selected 60 images from
the validation set using seed 42 and performed 8 untimed calls
to the pipeline before saving the outputs.
For inference latency, I used time.perf_counter( ) with
CUDA synchronization before and after each timed operation in
order to compensate for asynchronous execution by the GPU.

Al-A'areji et al. / Journal of Applied Science and Technology Trends Vol. 07, No. 02, pp. 534-553, (2026)

548
The times I reported include image decoding, preprocessing,
transferring one input tensor at a time, global classification,
predicted ROI probability, postprocessing and cropping of ROI
by the CPU, local classification, and confidence -adaptive
probability fusion in each respective pipeline. The timed steps
do not include loading the model checkpoint, initializing the
model, or initial model transfer to device.
The implementation used regular FP32 model parameters
but executed inferencing operations eligible for CUDA
automatic mixed precision. It did not perform any post-training
quantization, permanent FP16 weight conversion, or structured
pruning. The total inf erence pipeline involve three separate
networks all instantiation: the global classifier with 4,165,681
parameters, the ROI prediction network with 4,245,722
parameter, and local classifier with 4,165,681 parameters. The
total number of the pipeline is 12,577,084 parameters.
Based on an FP32 (single-precision float) parameter size of
4 bytes, we estimate the theoretical framework storage
requirement to be approximately 50.31 MB (47.98 MiB) for all
parameters only (excluding buffers, intermediate activations,
framework overhead , and checkpoint metadata, i.e. does not
include the size of the checkpoint file nor the runtime memory
footprint).
Peak GPU Memory GPU memory was measured using the
torch. cuda. Max memory allocated() on PyTorch (after some
warm-up iterations), with peak allocations reset prior to the start
of timed benchmark. This is a measurement of the maximum
amount of memory allocated through the PyTorch CUDA
allocator (including the resident model tensors and the evaluated
intermediate allocations). This does not include total device
memory, OS memory, or the peak application -level memory
measured on an embedded device.
The reported total end -to-end inference latency across all
local and global -local scales on Table V of 18.80/6.70 ms per
image (Straw Net -ROI/ global -only) and total 95 -percentile
latency of 19.24/7.03 ms/ image (Straw Net-ROI/ global-only),
respectively, for the configured batch-one size (the mean of all
the sequential runs) correspond to processing speeds of about
53.2/149.3 images/sec (Straw Net -ROI/ global -only). These
latencies are measured across all resolutions and scales (both
local and global).
TABLE V. INFERENCE COMPLEXITY AND COMPUTATIONAL
PERFORMANCE OF STRAWNET-ROI
Metric Global only Straw Net-ROI
Parameters (M) 4.17 12.58
Theoretical FP32 parameter storage
(MB)
16.66 50.31
Median latency (ms) 6.67 18.79
Mean latency (ms) 6.70 18.80
p95 latency (ms) 7.03 19.24
Batch-one throughput (images/s) 149.3 53.2
Peak PyTorch GPU memory
allocated (MiB)
43.36 77.23

Peak PyTorch-allocated GPU memory was 77.23 MiB for
the complete framework and 43.36 MiB for global -only
classification. Therefore, the complete framework requires
approximately 3.02 times as many parameters and 2.81 times the
mean inference latency of the global -only model under the
evaluated configuration.
These measurements provide a reproducible description of
the computational requirements on the tested server GPU.
However, they do not demonstrate real -time execution on
smartphones, embedded processors, or low -power edge
accelerators. No representative mo bile or embedded hardware
experiments were conducted, and device -level energy
consumption, thermal behavior, sustained inference
performance, and application -level peak memory were not
evaluated.
As a result, as Straw Net -ROI is not claimed to fulfill a
specific lightweight model -size requirement nor be optimized
for ready -to-implement, resource -constrained realizations.
Rather, the framework is evaluated as a double -view classifier,
which potentia lly incurs higher computational costs while
including the readily-located image data.
Future investigations may consider parameter sharing,
model compression, hardware -aware optimization, and
deployment-specific benchmarking. Claims of real-time mobile
or embedded operation will require measurements on
representative target devices under ex plicitly defined latency,
memory, and energy constraints.
TABLE VI. COMPUTATIONAL COST OF THE GLOBAL-ONLY
AND COMPLETE STRAWNET-ROI INFERENCE PIPELINES
Metric Global-Only StrawNet-ROI
Parameters (M) 4.17 12.58
Mean inference latency (ms) 6.70 18.80
95th-percentile latency (ms) 7.03 19.24
Throughput (images/s) 149.3 53.2
Peak allocated GPU memory (MiB) 43.4 77.2
Our end-to-end system showed about a 2.81 times the mean
inference latency as the global-only classifier. This appears to be
attributed to the additional time required for the frame -work to
perform in-room ROI prediction, by generating local view image
crops and performing local view classification. The total number
of parameters required was roughly 3x times that of the single -
view model. These measures provide some indication of
computational performance on the given GPU hardware for this
task, but do not necessarily demonstrate whether these
approaches are suitable for deployment on mobile or edge device
platforms. Note that there was no inference testing performed on
embedded CPU (x86) or mobile processor platforms, or on low-
power edge accelerators.
As a result, the framework does not claim to satisfy
constraints on latency, memory, and energy for mobile devices.
The classification accuracy achieved should be weighed against

Al-A'areji et al. / Journal of Applied Science and Technology Trends Vol. 07, No. 02, pp. 534-553, (2026)

549
the additional computational cost of the dual-view approach. In
follow-up work we expect that we will look into optimized
deployment, such as model compression, quantization, and
hardware-accelerated inference . Device -level benchmarking
will be necessary to determine whether the proposed framework
can operate within the resource and latency constraints of
practical smart-agriculture systems.
E. Comparative Analysis with State-of-the-Art Models
The results were compared with TinyViT -5M, Efficient
Former-L1, MobileViT-XXS, a teacher architecture based on
Efficient Net-B2, and the global-only version of Straw Net-ROI.
The baseline was chosen to include convolutional, mobile -
centric, and small -scale vision -transformer architectures with
different parameter budgets.
All reported classification models were evaluated on the
same duplicate -audited dataset partition containing 1,426
training, 294 validation, and 737 test images. Five fixed random
seeds (42–46) were used, and the principal evaluation metrics
were accuracy, macro-precision, macro -recall, and macro -F1.
The evaluation summaries published include the mean and the
sample standard deviation of five experimental runs. Baseline
implementation and pretrained initialization methods were used
for the experiment. The three models TinyViT -5M, Efficient
Former-L1 an d MobileViT -XXS, were implemented by the
model library of the timm model, tinyvit5m224.in1k,
efficientformerl1, and mobilevit_xxs, respectively. Also, these
models were pretrained weights, and the number of classes of
classification head was set to the seven classes . The
EfficientNet-B2-based teacher was a separate multiscale
architecture employing an EfficientNet-B2 feature encoder with
additional classification components; it was not a plain
EfficientNet-B2 classifier.
The three timm baselines shared a documented supervised
training recipe using AdamW , a nominal learning rate of 3 ×
10⁻⁴, weight decay of 10⁻⁴, batch size 32, and a maximum of 100
training epochs with early -stopping patience of 30. Their
optimization used a five -epoch learning-rate warmup followed
by cosine decay, together with exponenti al moving average
checkpoint handling and class -weighted focal cross -entropy.
The EfficientNet-B2-based teacher followed a separate training
configuration with a maximum of 80 epochs and patience of 25.
All baseline architectures were evaluated using 224 × 224
images. The global-only classifier and complete StrawNet-ROI
framework used 288 × 288 images. Moreover, the proposed
framework required sequential global-classifier training, spatial
ROI supervision, and local -classifier fine-tuning, followed by
validation-based selection of a confidence -adaptive fusion
policy.
Comparability and optimization limitations. Although the
models shared the audited dataset partition, evaluation metrics,
and repeated-run reporting scheme, the comparison was not a
strictly controlled identical -resolution or equal -compute
benchmark. The pretrained initialization methods, architecture-
specific optimization procedures, and training schedules
differed. A matched hyperparameter -search budget was not
established for all evaluated architectures. Accordingly, the
results should not be interpreted as isolating architectural quality
independently of input resolution, initialization, or training
resources.
Model complexity. Parameter counts were obtained from
the implemented architectures. Theoretical FP32 parameter
storage was calculated using four bytes per parameter. These
storage estimates exclude nonparameter buffers, serialized
checkpoint metadata, int ermediate activations, and runtime
framework overhead, and therefore do not represent measured
application-level memory requirements.
Inference benchmarking limitations. Historical baseline
latency measurements were collected using batch-one, forward-
only inference on synthetic input tensors, following warm -up
and CUDA synchronization. By contrast, the reported StrawNet-
ROI benchmark measured the complete image -processing and
dual-view inference pipeline, including image decoding,
preprocessing, ROI extraction, classification, and probability
fusion. Because these measurements do not use an identical
timing scope and wer e not established as a matched hardware
benchmark, forward -only baseline latencies are not directly
compared with the complete StrawNet-ROI end-to-end latency.
Interpretation of comparative results. As presented in Table
VI, StrawNet-ROI obtained a mean test macro -F1 of 97.79 ±
0.47%, compared with 97.53 ± 0.59% for TinyViT -5M and
97.50 ± 0.77% for EfficientFormer -L1. These differences are
numerically positive but relatively small. They do not establish
statistically significant superiority over reference architectures.
TABLE VII. COMPARISON OF STRAWNET-ROI WITH LIGHTWEIGHT AND STATE-OF-THE-ART CLASSIFICATION ARCHITECTURES
Model Test Accuracy (%) Test Macro-F1 (%) Parameters (M) Input Resolution
StrawNet-ROI 98.78 ± 0.25 97.79 ± 0.47 12.58 288 × 288
TinyViT-5M 98.53 ± 0.28 97.53 ± 0.59 5.07 224 × 224
EfficientFormer-L1 98.67 ± 0.35 97.50 ± 0.77 11.40 224 × 224
EfficientNet-B2 98.59 ± 0.37 97.21 ± 0.72 7.53 224 × 224
Global-only classifier 98.56 ± 0.28 97.20 ± 0.60 4.17 288 × 288
MobileViT-XXS 98.24 ± 0.19 96.48 ± 0.60 0.95 224 × 224

Al-A'areji et al. / Journal of Applied Science and Technology Trends Vol. 07, No. 02, pp. xx –xx, (2026)

550
The reuse of the internal test partition throughout history
and the lack of an independently collected external evaluation
dataset further may imply that generalization strength of the
claims made were weaker than claimed. The results presented
here sugges t no more than competitive exploratory internal
classification results and not, at least not alone, superior
predictive accuracy, computational efficiency, or deployment
readiness.
F. Ablation Study
To understand the relative contributions of the high -level
architectural and decision-level components of Straw Net-ROI,
we performed a quantitative ablation study. The evaluation was
two-pronged: the global classifier network was subjected to
component-removal experiments, and different algorithms for
the local-view classifier and local -view prediction fusion were
compared. Each experiment was run with five different random
seeds (42-46). Because of different data partitions and testing
methodologies, these experiments are presented separately.

  1. Architectural Component Ablation
    The first experiment explored the effects of removing
    feature pyramid fusion (FPN), lightweight channel -spatial
    attention (LCSA), and scale-aware gating (SAG) from the global
    classifier. For the experiments, the same test split of 737 images
    at 288 × 288 pixels, input resolution, was used to evaluate all the
    setups.
    TABLE VIII. QUANTITATIVE ABLATION OF GLOBAL CLASSIFIER COMPONENTS
    Configuration Accuracy (%) Macro-Precision (%) Macro-Recall (%) Macro-F1 (%) Parameters (M) FP32 Storage (MB)
    Full Global Classifier 98.56 97.69 96.82 97.20 ± 0.60 4.166 16.66
    Without FPN 98.43 97.24 96.30 96.70 ± 0.61 4.010 16.04
    Without LCSA 98.64 97.85 97.01 97.37 ± 0.88 4.166 16.66
    Without Scale Gating 98.59 97.65 97.03 97.28 ± 0.81 4.145 16.58
    Results are given as five -run averages, macro -F1 scores
    with mean ± sample standard deviation in percentage points.
    Experimental procedure: Components were removed with
    limited fine -tuning of pretrained configurations, rather than
    independently trained fro m random initializers. Removing the
    Feature Pyramid Network (FPN): When the FPN feature
    extractor was absent, the mean macro-F1 score was lowered by
    ≈0.50 percentage points, indicating possible influence from
    multi-scale feature aggregation, but exploratory paired statistical
    test did not reach the nominal threshold (p = 0.092).
    On the other hand, the cancellation of LCSA probably
    brought the macro-F1 to 97.37%, while the cancellation of scale-
    aware gating probably brought it to 97.28%. However, these
    differences were statistically insignificant. These results, in turn,
    lend evidence that independent efficiency improvements from
    the attention module and the gating module are not evidenced by
    the experimental protocol adopted in this work. These are,
    therefore, kept as components of the global architecture, rather
    than as performance refinements tested independently.
  2. Local-View and Prediction-Fusion Ablation
    The second experiment examined the effects of ROI-based
    cropping and global -local prediction fusion. This analysis was
    performed on the validation split of 294 images with the same
    five random seeds. It examined global-only classification versus
    independently-trained classifiers for local crops and confidence
    adaptive fusion of global and local prediction. Additionally, to
    determine whether ROI -based cropping potentially provides
    biologically relevant spatial information, two further local-view
    controls were tested with center crops and random crops, which
    used the predicted dimensions of the associated ROI as an
    estimate to compare the location of crops, while controlling for
    crop size.
    TABLE IX. QUANTITATIVE ABLATION OF LOCAL-VIEW SELECTION AND PREDICTION FUSION
    Configuration Accuracy (%) Macro-Precision (%) Macro-Recall (%) Macro-F1 (%)
    Global-only classification 98.71 97.28 98.22 97.70 ± 0.57
    ROI crop-only 98.03 96.78 96.44 96.54 ± 1.04
    Center crop-only 98.78 97.89 97.81 97.82 ± 0.34
    Random crop-only 97.07 95.40 95.23 95.23 ± 1.01
    Fixed fusion (w = 0.25) 98.71 97.28 98.22 97.70 ± 0.57
    Fixed fusion (w = 0.50) 98.84 97.75 98.31 98.00 ± 0.68
    Fixed fusion (w = 0.75) 98.30 97.20 96.83 96.94 ± 1.16
    ROI + confidence-adaptive fusion (α = 0.30, β = 5.0) 98.91 97.80 98.37 98.05 ± 0.78
    Center + confidence-adaptive fusion 98.78 97.40 98.08 97.71 ± 0.52
    Random + confidence-adaptive fusion 98.44 96.80 97.32 97.03 ± 0.63

Al-A'areji et al. / Journal of Applied Science and Technology Trends Vol. 07, No. 02, pp. xx –xx, (2026)

551
All results were sampled from the validation partition; all
results are from means of five runs and reported as macro-F1 as
(mean ± sample standard deviation). The center - and random-
crop controls were size-matched to crops based on ROI and not
from an ent irely separate crop sizing policy. The center -crop
classifier trained only on ROI crops returned a mean macro -F1
of 96.54%, relative to 97.70% for the global -only classifier.
More surprisingly, classification using only center crops was
97.82%, outperforming the (ROI crop only) classifier to center
crops. This suggests that cropping based solely on whether a
region was deemed ROI may not necessarily provide the most
robust discriminative representation of the data.
The combination of global and region of interest (ROI) -
based local predictions resulted in the highest mean validation
macro-F1 obtained (98.05%). The fixed -weight fusion
configuration, which maintained equal global and local
contributions, scored 98.00%, and the selected confidence -
adaptive configuration scored 98.05%. Therefore, the numerical
gain in improvement with confidence -adaptive weighting was
small. The confidence -fusion configuration based on the ROI
was superior to the global -only classification by 0.35% and to
the centre-crop fusion by 0.34%. However, the paired
comparisons were not statistically significant (p = 0.075 and p =
0.457). None of the four proposed ROI -versus-control tests
remained statistically significant after adjusting for multiple
testing with the Holm correction.
These results suggest that combining complementary global
and local predictions may be useful, but they do not establish
that ROI -based spatial selection is consistently superior to
conventional cropping alternatives. The confidence -fusion
policy was originally selected using validation performance for
the ROI configuration; consequently, the subsequent center- and
random-crop comparisons should also be interpreted as
exploratory.
3. Training Objective and Distillation Considerations
The training objectives of StrawNet -ROI differ across its
constituent networks. The global classification model was
trained using class -weighted focal cross -entropy with a focal
parameter of 1.5, label smoothing of 0.05, and frequency -
derived class weights proportional to the inverse square root of
the training -class frequencies. The ROI prediction head was
trained using a combination of binary cross-entropy with logits
and soft Dice loss. The local classification network was
subsequently fine -tuned using class -weighted cross -entropy
with label smoothing of 0.025 rather than a focal-loss objective.
Although class-dependent weighting was incorporated into
the training procedure, its independent contribution to minority-
class recognition was not established through a matched loss -
function ablation for the complete StrawNet -ROI framework.
Therefore, the relatively high class -wise scores and the
numerical improvements observed for selected minority classes
should not be interpreted as causal evidence of improved
performance specifically attributable to focal loss or class
weighting.
Furthermore, the final framework does not use Teacher –
Assistant–Student knowledge distillation, an intermediate
assistant model, or distillation -based optimization. These
components are not claimed as contributions of the evaluated
architecture.
Accordingly, the reported class-wise results characterize the
observed behavior of the complete dual-view framework under
the audited internal evaluation protocol, while the effects of
individual loss-function components remain unverified.
These architectural ablations were done by removing
subcomponents and doing limited finetuning on pretrained
architectures rather than training each combination of
components from scratch with the same initializations. As such
these experiments are a rough indication of component
sensitivity, but do not establish the independent and/or combined
importance of all architectural components.
Although class-weighted focal loss and exponential moving
average (EMA) were employed during global-classifier training,
their individual effects were not isolated through matched
ablation experiments. EMA was used for training stabilization
and checkpoint selection rather than as a separately validated
methodological contribution.
4. Overall Interpretation
The ablation results provide mixed evidence regarding the
contributions of individual architectural components. Multi -
scale fusion showed a positive descriptive effect on the global
classifier, whereas the attention and scale -gating modules did
not exhibit consistent independent benefits. At the system level,
confidence-adaptive global–local fusion achieved the highest
mean validation macro -F1, although its advantage over
controlled cropping alternatives was not statistically conclusive.
These findings support a restrained interpretation of the
proposed framework as a dual -view classification system with
competitive internal performance and additional computational
overhead, rather than as a collection of components each proven
to improve classification. The component -removal protocol,
repeated use of the historical test set during development, and
limited statistical power are acknowledged as constraints on
causal and generalization claims.
V. CONCLUSION
This study presented StrawNet -ROI, an annotation-guided
dual-view framework for seven -class strawberry disease
classification. The proposed system combines a multiscale
global classifier, an automatically supervised ROI prediction
network, spatially constrained local -image extraction, and
confidence-adaptive prediction fusion. Experiments conducted
on a duplicate -audited dataset of 2,457 images across five
random seeds achieved a mean internal test accuracy of 98.78 ±
0.25% and a macro-F1 score of 97.79 ± 0.47%. Compared with
the global-only classifier, StrawNet-ROI achieved a numerical
macro-F1 improvement of 0.59 percentage points. However, the
paired comparison was not statistically significant (p = 0.1268).
Additional cropping controls and architectural ablations
likewise did not establish consistent, statistically significant

Al-A'areji et al. / Journal of Applied Science and Technology Trends Vol. 07, No. 02, pp. 534-553, (2026)

552
benefits for all individual components. The results suggest that
integrating global and automatically selected local
representations may provide complementary information for
strawberry disease classification. However, the combined
framework required 12.58 million inference parameters and
achieved an average end-to-end latency of 18.80 ms on a server-
class GPU which is a significant computational overhead
compared to the global-only model. Several limitations restrict
the interpretation of the above results. These were obtained from
a retrospectively audited internal dataset, with historic al test
images that were already reviewed during the development of
the methodology. Moreover, independence at the plant level and
acquisition-session could not be established with certainty, nor
was any independent external seven-class evaluation performed.
Therefore, no conclusions regarding robustness to unseen field
conditions nor suitability for deployment in resource -
constrained environments are made herein. We recommend
future studies investigate independent multi-location validation,
further annotat ion improvements, more rigorous localiser
evaluation, and head -to-head comparisons with fixed training
budgets. Parameter sharing, model compression and hardware -
specific benchmarking may also reveal the utility of the
framework in the embedded setting.
ACKNOWLEDGMENT
We would like to thank Ninevah University for supporting this work.
DATASET AVAILABILITY
The Strawberry Disease Detection Dataset, available at
https://www.kaggle.com/datasets/usmanafzaal/strawberry-
disease-detection-dataset, is described in terms of conditions
under which data was obtained and original annotation process
by Afzaal et al. (2021) [4]. In addition, our authors retained the
tables of audited split manifest, duplicate -screening logs,
preprocessing parameters, and experiment logs that are
necessary for reproduction. These secondary files have not been
verified to be publicly deposited supplementary information.
REFERENCES
[1] S. Savary, L. Willocquet, S. J. Pethybridge, P. Esker, N. McRoberts,
and A. Nelson, “The global burden of pathogens and pests on major
food crops,” Nature Ecology & Evolution 2019 3:3, vol. 3, no. 3, pp.
430–439, Feb. 2019, doi: 10.1038/s41559-018-0793-y.
[2] A. K. Mahlein, “Plant disease detection by imaging sensors –
Parallels and specific demands for precision agriculture and plant
phenotyping,” Plant Dis., vol. 100, no. 2, pp. 241 –254, Feb. 2016,
doi:10.1094/PDIS-03-15-0340-
FE;WGROUP:STRING:PUBLICATION.
[3] S. Zhao, J. Liu, and S. Wu, “Multiple disease detection method for
greenhouse-cultivated strawberry based on multiscale feature fusion
Faster R_CNN,” Comput. Electron. Agric., vol. 199, p. 107176, Aug.
2022, doi: 10.1016/J.COMPAG.2022.107176.
[4] U. Afzaal, B. Bhattarai, Y. R. Pandeya, and J. Lee, “An Instance
Segmentation Model for Strawberry Diseases Based on Mask R -
CNN,” Sensors 2021, Vol. 21, Page 6565 , vol. 21, no. 19, p. 6565,
Sep. 2021, doi: 10.3390/S21196565.
[5] S. P. Mohanty, D. P. Hughes, and M. Salathé, “Using deep learning
for image-based plant disease detection,” Front. Plant Sci., vol. 7, no.
September, p. 215232, Sep. 2016, doi:
10.3389/FPLS.2016.01419/TEXT.
[6] K. P. Ferentinos, “Deep learning models for plant disease detection
and diagnosis,” Comput. Electron. Agric., vol. 145, pp. 311–318, Feb.
2018, doi: 10.1016/J.COMPAG.2018.01.009.
[7] E. C. Too, L. Yujian, S. Njuki, and L. Yingchun, “A comparative
study of fine -tuning deep learning models for plant disease
identification,” Comput. Electron. Agric., vol. 161, pp. 272–279, Jun.
2019, doi: 10.1016/J.COMPAG.2018.03.032.
[8] J. G. A. Barbedo, “Factors influencing the use of deep learning for
plant disease recognition,” Biosyst. Eng., vol. 172, pp. 84–91, Aug.
2018, doi: 10.1016/J.BIOSYSTEMSENG.2018.0

论文原文与出版信息

Journal of Applied Science and Technology Trends
查看论文原文

继续探索