基于改进YOLO 11 的梨园复杂环境下梨果检测方法

Pear Detection Method in Complex Orchard Environments Based on Improved YOLO 11

摘要:

针对梨园环境下梨果检测易受光照变化、目标尺度差异、分布密集以及严重遮挡等复杂因素影响的问题，本文提出一种基于改进YOLO 11的梨果检测方法。该方法在YOLO 11基础网络中引入自监督形状感知特征增强（Self-supervised shape-aware feature enhancement，SS-SAFE）模块，通过自监督学习预测目标长宽比，从而增强对梨果独特形状特征的感知与提取能力;同时，采用自适应核卷积（Adaptive kernel convolution，AKConv）替换原始C3K2（Cross stage partial network with kernel size 2）模块中的标准卷积，构建C3K2-AKConv（Cross stage partial network with kernel size 2 and adaptive kernel convolution）深层特征提取结构，提升网络对不同尺度梨果的适应性;此外，将C2PSA（Convolutional block with parallel spatial attention）模块替换为C2PSA-HSAN（Cross-stage partial with hybrid spectral attention network）模块，通过优化特征交互与融合机制，可显著增强模型在复杂环境下的特征融合能力。试验结果表明，改进模型在精确率、召回率、F1值、mAP@0.5和mAP@0.5:0.95分别达到88.9%、81.7%、85.2%、88.6%和67.2%，相较于YOLO 11n模型分别提升3.8、1.9、2.7、3.2、2.9个百分点。研究结果可为梨园智能化管理与自动化采摘提供技术支撑。

Abstract:

Aiming to address the challenges of detecting pear fruits in orchard environments, such as variable illumination, multi-scale targets, dense distributions, and severe occlusions, a pear detection method was proposed based on an improved YOLO 11 framework. The proposed approach integrated a self-supervised shape-aware feature enhancement (SS-SAFE) module into the base YOLO 11 network, which learned to predict the aspect ratio of targets through self-supervised learning, thereby strengthening the perception and extraction of distinctive pear-shape features. Additionally, an adaptive kernel convolution (AKConv) was used to replace the standard convolution in the original cross stage partial network with kernel size 2 (C3K2) module, forming a cross stage partial network with kernel size 2 and adaptive kernel convolution (C3K2-AKConv) deep feature extraction structure, improving the network's adaptability to pears of varying scales. Furthermore, the C2PSA (convolutional block with parallel spatial attention) module was replaced by the C2PSA-HSAN (cross-stage partial with hybrid spectral attention network) module, which optimizes the feature interaction and fusion mechanism, significantly enhancing the model's feature fusion ability in complex environments. Experimental results show that the improved model achieves a precision of 88.9%, recall of 81.7%, F1-score of 85.2%, mAP@0.5 of 88.6%, and mAP@0.5:0.95 of 67.2%, outperforming the baseline YOLO 11n by 3.8, 1.9, 2.7, 3.2, 2.9 percentage points, respectively. This study offers a technical reference for the intelligent management and automated harvesting in pear orchards.