基于改进YOLO v5的复杂环境下桑树枝干识别定位方法

为实现复杂自然环境下对桑树嫩叶处枝干的识别检测，改变当前桑叶采摘设备作业过程中依赖人工辅助定位的现状，解决识别目标姿态多样和环境复杂导致的低识别率问题，提出一种基于改进YOLO v5模型的桑树枝干识别模型(YOLO v5-mulberry),并结合深度相机构建定位系统。首先，在YOLO v5的骨干网络中加入CBAM(Convolutional block attention module)注意力机制，提高神经网络对桑树枝干的关注度；并增加小目标层使模型可检测4像素×4像素的目标，提高了模型检测小目标的性能；同时使用GIoU损失函数替换原始网络中的IoU损失函数，有效防止了预测框和真实框尺寸较小时无法正确反映预测框及真实框之间位置关系的情况；随后，完成深度图和彩色图的像素对齐，通过坐标系转换获取桑树枝干三维坐标。试验结果表明：YOLO v5-mulberry检测模型的平均精度均值为94.2%,较原模型提高16.9个百分点，置信度也提高12.1%;模型室外检测时应检测目标数53,实际检测目标数为48,检测率为90.57%;桑树嫩叶处枝干三维坐标识别定位系统的定位误差为(9.498 5 mm...

Mulberry Branch Identification and Location Method Based on Improved YOLO v5 in Complex Environment

In order to solve the recognition and detection of branches at the young leaves of mulberry trees in complex natural environments, overcome the current situation of relying on manual assisted positioning in the operation process of mulberry leaf harvesting equipment, and improve the problem of low recognition rate caused by diverse target postures and complex environments, a mulberry branch and trunk recognition model was proposed based on the improved YOLO v5 model (YOLO v5-mulberry) and combined it with the depth camera to construct a location system. Firstly, convolutional block attention module (CBAM) attention mechanism was added to the backbone network of YOLO v5 to improve the neural network’s attention to the mulberry branches;and a small target layer was added to enable the model to detect 4pixels×4pixels targets, which improved the model’s performance in detecting small targets. At the same time, the GIoU loss function was used to replace the IoU loss function in the original network, which effectively prevented the position relationship between the prediction box and the real box from being correctly reflected when the size of the prediction box and the real box was small. Subsequently, the pixel alignment of the depth map and the color map was completed, and the 3D coordinates of the mulberry tree trunk were obtained through the conversion of the coordinate system. The test results showed that the average accuracy of YOLO v5-mulberry detection model was 94.2%, which was 16.9 percentage points higher than that of the original model, and the confidence level was also 12.1% higher;the number of targets that should be detected by the model outdoor detection was 53, and the number of actually detected targets was 48, and the detection efficiency was 90.57%;the positioning error of the three-dimensional coordinate recognition and location system of the mulberry branch and trunk at the tender leaves was (9.4985mm,11.285mm,19.11mm), which met the requirements for use. The research result can achieve the recognition and positioning of branches and trunks at the tender leaves of mulberry trees, which can help to further promote the research, development and application of intelligent mulberry leaf picking robots.