基于改进YOLO v5的复杂环境下花椒簇识别与定位方法

花椒树产果量大，枝干纵横交错，树叶茂密，给花椒的自动化采摘带来了困难。因此，本文设计一种基于改进YOLO v5的复杂环境下花椒簇的快速识别与定位方法。通过在主干提取网络CSPDarknet的CSPLayer层和Neck的上采样之后增加高效通道注意力ECA(Efficient channel attention)来简化CSPLayer层的计算量，提升了特征提取能力。同时在下采样层增加协同注意力机制CA(Coordinate attention),减少下采样过程中信息的损失，强化特征空间信息，配合热力图(Grad-CAM)和点云深度图，来完成花椒簇的空间定位。测试结果表明，与原YOLO v5相比较，改进的网络将残差计算减少至1次，保证了模型轻量化，提升了效率。同帧数区间下，改进后的网络精度为96.27%,对比3个同类特征提取网络YOLO v5、YOLO v5-tiny、Faster R-CNN,改进后网络精确度P分别提升5.37、3.35、15.37个百分点，连株花椒簇的分离识别能力也有较大提升。实验结果表明，自然环境下系统平均识别率为81.60%、漏检率为18.39%,能够满足花椒簇识别...

Recognition and Localization Method for Pepper Clusters in Complex Environments Based on Improved YOLO v5

Pepper trees yield is a substantial quantity of fruits, characterized by crisscrossed branches and dense foliage, resulting insignificant challenges for automated peppercorn picking. Therefore, a fast identification and localization method of pepper clusters in complex environment based on improved YOLO v5 was proposed. By adding efficient channel attention (ECA) after the CSPLayer of the backbone extraction network CSPDarknet and the upsampling layer of Neck to simplify the computation of the CSPLayer layer and improve the feature extraction capability. In the downsampling layer, coordinate attention (CA) was added to reduce the loss of information in the downsampling process, strengthen the spatial information of features, and cooperate with the heat map (Grad-CAM) and the depth map of the point cloud to complete the spatial localization of pepper clusters. The test results showed that the improved network over the original YOLO v5 reduced the residual computation to 1 time, which ensured the model was lightweight and the efficiency was improved. Under the same frame number interval, the accuracy of the improved network was 96.27%, comparing with three similar feature extraction networks YOLO v5, YOLO v5-tiny, and Faster R-CNN, the precision of the improved network was improved by 5.37 percentage points, 3.35 percentage points, and 15.37 percentage points, respectively, and the ability of separating and recognizing the pepper clusters of the successive plants was greatly improved. The experimental results showed that the average checking accuracy of the system in the natural environment was 81.60%, and the leakage rate was 18.39%, which can satisfy the pepper cluster recognition, and build the foundation for mobile deployment.