基于改进YOLO v5的桑叶采摘与桑枝伐条识别定位方法

Method for Detection and Localization of Mulberry Leaf Harvesting and Branch Pruning Based on Improved YOLO v5

摘要:

为应对桑树生长的多季节变化及形态多样性，同时满足其与养蚕周期相匹配的需求，解决当前桑叶自下而上采摘与桑枝伐条作业中对人工定位的依赖问题，本研究建立桑园2021年7月、2024年9月和11月各种气候与桑枝形态的数据集，提出基于改进YOLO v5的桑树枝干检测模型YOLO v5-cytp，构建基于深度相机的三维定位系统实现精准识别。首先，加入CA注意力机制，增强模型对桑枝底部的特征聚焦能力；然后将YOLO v5基础的CIoU损失函数替换为SIoU损失函数以提升训练速度与推理精度；最后采用轻量化GhostNet重构YOLO v5的骨干网络，在满足使用要求的前提下将模型轻量化。完成深度相机标定，实现RGB图像与深度图像对齐，经过坐标转换最终获取目标三维坐标。试验表明，YOLO v5-cytp模型平均精度均值达93.4%，相较YOLO v5原始模型提高1.2个百分点；同时内存占用量由3.79MB降低为3.02MB，降低20.31%；模型桑枝识别率达到91.11%；桑枝底部三维坐标（X，Y，Z）定位最大误差为（11.3，14.1，27.0）mm，符合误差允许值。模型可同时实现桑叶自下而上采摘与桑枝伐条作业的识别定位，可为桑园智能化采摘与伐条机器人提供参考。

Abstract:

Aiming to address the multi-seasonal growth and morphological diversity of mulberry trees, and meet the demands of synchronizing with the silkworm breeding cycle, a dataset encompassing various climatic conditions and mulberry branch morphologies in July 2021, September 2024, and November 2024 was established. An improved YOLO v5-based mulberry branch detection model, YOLO v5-cytp was proposed, and a 3D positioning system was developed by using a depth camera to achieve precise identification. Firstly, the CA attention mechanism was incorporated to enhance the model’s feature extraction capability for the lower parts of mulberry branches. Secondly, the original CIoU loss function in YOLO v5 was replaced with the SIoU loss function to improve training speed and inference accuracy. Finally, the lightweight GhostNet was adopted to reconstruct the backbone network of YOLO v5, ensuring the models performance while reducing its size. The calibration of the depth camera was completed, alignment between RGB images and depth images was achieved, and the target 3D coordinates were obtained through coordinate transformation. Experimental results showed that the YOLO v5-cytp model achieved an average precision of 93.4%, representing a 1.2 percentage points improvement over the original YOLO v5 model. Meanwhile, the memory footprint was reduced from 3.79MB to 3.02MB, with a reduction of 20.31%. The identification rate of the model for mulberry branches reached 91.11%, and the maximum errors in the 3D coordinate positioning of the lower part of the branches (X, Y, Z) were (11.3, 14.1,27.0)mm, respectively, which were within the allowable error range. The research result can simultaneously realize the identification and positioning of mulberry leaf picking from bottom to top and branch pruning operations, providing a reference for the development of intelligent mulberry harvesting and pruning robots.