基于MEW‑YOLO的库尔勒香梨目标检测方法

Object Detection Method for Korla Fragrant Pears Based on MEW‑YOLO

摘要:

库尔勒香梨果实检测是实现智能化低损采摘的关键环节，其检测精度与实时性直接影响采摘机器人的作业效率与稳定性。然而，受制于果实尺寸差异、光照变化、枝叶遮挡及复杂背景等因素，现有检测方法在实际应用中仍面临挑战。为提高库尔勒香梨的识别准确性与作业适应性，本文提出一种基于YOLO 11n的轻量化改进检测模型MEW?YOLO，通过引入混合局部通道注意力模块（Mixed local channel attention，MLCA）、高效上卷积模块（Efficient up?convolution block，EUCB）以及加权交并比损失函数（Wise intersection over union，WIoU），分别从特征增强、空间细节恢复与边界回归优化3个层面提升模型对复杂场景的适应性。其中，MLCA融合局部与全局通道信息以增强多尺度特征表达;EUCB在特征融合阶段强化空间信息重建以提升小目标与遮挡目标表征;WIoU通过样本质量感知加权与聚焦调节优化回归梯度分配，增强遮挡与重叠条件下的定位鲁棒性。在采集的库尔勒香梨自然场景数据集上，MEW?YOLO相较基准YOLO 11n的mAP@0.5、mAP@0.5:0.95、精确率与召回率分别提升2.7、3.3、5.3、3.7个百分点;参数量仅增加约3.4%，浮点运算量为6.8×10^9。研究结果表明，MEW?YOLO可为库尔勒香梨自然果园场景下的自动化检测与后续采摘作业提供可靠的视觉输入，并为特色水果目标检测模型的轻量化设计提供参考。

Abstract:

Accurate detection of Korla fragrant pear fruit is a key prerequisite for intelligent low?damage harvesting, as detection accuracy and real?time performance directly affect the operational efficiency and stability of harvesting robots. However, existing methods still face challenges in practical orchard applications due to variations in fruit size, illumination changes, occlusion by leaves and branches, and complex backgrounds. To improve detection accuracy and operational adaptability for Korla fragrant pear, a lightweight improved detection model, termed MEW?YOLO, was proposed based on YOLO 11n. MEW?YOLO introduced a mixed local channel attention (MLCA) module, an efficient up?convolution block (EUCB), and a wise intersection over union (WIoU) loss function, thereby enhancing model adaptability to complex scenes from three aspects: feature enhancement, spatial detail restoration, and bounding?box regression optimization. Specifically, MLCA integrated local and global channel information to strengthen multi?scale feature representation;EUCB reinforced spatial information reconstruction during feature fusion to improve the representation of small and occluded targets;and WIoU optimized the allocation of regression gradients through sample?quality?aware weighting and focusing regulation, enhancing localization robustness under occlusion and overlap conditions. On the collected Korla fragrant pear dataset captured under natural orchard conditions, MEW?YOLO outperformed the baseline YOLO 11n by 2.7, 3.3, 5.3, and 3.7 percentage points in mAP@0.5, mAP@0.5:0.95, precision, and recall, respectively;it used 2.67 ×10^6 parameters and required 6.8×10^9 of computation. The results indicated that MEW?YOLO can provide reliable visual input for automated detection and subsequent harvesting operations in natural Korla fragrant pear orchards, and offered a reference for the lightweight design of detection models for specialty fruit targets.