[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"item-2499":3},{"id":4,"title":5,"url":6,"summary":7,"summary_zh":8,"content":9,"source_name":10,"source_url":6,"published_at":11,"category":12,"cover_url":9,"hotness":13,"is_selected":14,"score":15,"score_detail":16,"sources":24,"tags":26,"view_count":32,"doi":33,"paper":34,"created_at":60},2499,"Real-time on-tree Korla pear grading with a shared ordinal–anomaly descriptor field","https:\u002F\u002Fdoi.org\u002F10.1016\u002Fj.compag.2026.112371","On-tree grading of Korla fragrant pears couples two decisions of different nature: a gradual Grade-A\u002FGrade-B appearance transition and a Grade-C surface-defect judgement, both made under illumination change, occlusion and scale variation. Detectors that treat the fruit as a single class, and graders that place the three grades on one flat classification axis, optimise a small A\u002FB deviation and an A\u002FC defect error through the same geometry and leave no explicit account of what ordinal and defect information each spatial location should preserve. We address this by adding an explicit ordinal–anomaly descriptor field to the shared feature hierarchy. A shallow module generates three channels – ordinal tendency, defect tendency and foreground confidence – that condition four downstream detector stages, and a hierarchical head converts the two kinds of evidence into normalised grade probabilities. On the orchard-disjoint KPR-3 dataset (3826 images, 12,458 instances, three commercial orchards), the model reaches 94.0% mAP@50 and 92.6% matched-instance grade accuracy, and lowers the unsafe-pick rate from 4.8% to 2.1%. On a Jetson Orin Nano at 15 W, the INT8 engine runs at 93.1% mAP@50 and sustains 42 FPS of end-to-end pipelined throughput. Monocular sensing remains a key limitation under severe front–back fruit overlap.","库尔勒香梨的树上分级耦合了两个性质不同的决策：渐变的A级\u002FB级外观过渡，以及C级表面缺陷判定，二者均在光照变化、遮挡和尺度变化下做出。将果实视为单一类别的检测器，以及将三个等级置于同一扁平分类轴上的分级器，通过相同的几何结构优化较小的A\u002FB偏差和A\u002FC缺陷误差，却未明确说明每个空间位置应保留何种序数信息和缺陷信息。我们通过在共享特征层级中增加一个显式的序数–异常描述符场来解决这一问题。一个浅层模块生成三个通道——序数倾向、缺陷倾向和前景置信度——用以调节四个下游检测器阶段，一个层级式头部将两类证据转换为归一化等级概率。在果园不相交的KPR-3数据集（3826幅图像、12458个实例、三个商业果园）上，模型达到94.0% mAP@50和92.6%的匹配实例等级准确率，并将不安全采摘率从4.8%降至2.1%。在15 W的Jetson Orin Nano上，INT8引擎以93.1% mAP@50运行，并维持42 FPS的端到端流水线吞吐量。在严重前后果实重叠情况下，单目感知仍是主要局限。",null,"Computers and Electronics in Agriculture","2026-09-14T00:00:00Z","论文",10,false,83,{"impact":17,"substance":18,"depth":19,"authority":20,"freshness":21,"relevant":22,"comment":23},18,23,19,14,9,1,"提出序数-异常描述子场的树上香梨实时分级方法，在自建KPR-3数据集与Jetson边缘端验证，方法新颖、数据扎实，对果园采摘机器人有直接参考价值。",[25],{"name":10,"url":6},[27,28,29,30,31],"智慧农业","农业人工智能","边缘计算","库尔勒香梨","水果分级",0,"10.1016\u002Fj.compag.2026.112371",{"doi":33,"openalex_id":35,"authors":36,"venue":10,"cited_by_count":32,"oa_url":6,"card":53,"direction":57,"ingested_from":59},"W7212908812",[37,39,42,44,46,48,50],{"name":38,"orcid":9},"Bingyu Cao",{"name":40,"orcid":41},"Peng Zhou","https:\u002F\u002Forcid.org\u002F0000-0002-6345-5307",{"name":43,"orcid":9},"Zhikai Yang",{"name":45,"orcid":9},"Wei Chen",{"name":47,"orcid":9},"Yingchao Wang",{"name":49,"orcid":9},"Mingqi Kan",{"name":51,"orcid":52},"Haiyong Chen","https:\u002F\u002Forcid.org\u002F0000-0002-5262-4208",{"tldr":54,"method":55,"finding":56,"direction":57,"opportunity":58},"提出共享序数-异常描述子场，实现树上库尔勒香梨实时分级。","在共享特征层级加入序数、缺陷、前景三通道描述子，用KPR-3数据集训练。","mAP@50达94.0%，分级准确率92.6%，不安全采摘率从4.8%降至2.1%。","智慧农业 \u002F 农业物联网","可探索多模态或深度传感融合，解决严重前后遮挡下的单目感知局限。","openalex","2026-09-15T23:30:01.440708Z"]