Abstract: To develop the function of a selfdriving soybean phenotypic information collection platform, achieve automated leaf recognition phenotypic parameters acquisition, and solve the problem of manual intervention, this study proposes a soybean canopy leaf recognition method based on improved YOLOv5s. The Convolutional Block Attention Module （CBAM） is introduced into the feature extraction layer of the backbone network to improve the recognition of small leaves. The Spatial Pyramid Pooling Fast （SPPF） structure is replaced with the Atrous Spatial Pyramid Pooling （ASPP） structure to enhance local information detection. The SIoU loss function is incorporated to improve the localization of leaves in complex backgrounds. The leaf length, leaf width, and leaf area are obtained by segmenting individual leaves using the coordinates of the bounding boxes. The experimental results show that the improved model outperforms the original model with a 5.4%, 3.3%, and 4.6% increase in mAP, precision, and recall, respectively. The average absolute errors for leaf length, leaf width, and leaf area are 0.98 cm, 0.56 cm, and 6.20 cm2, respectively.

Key words: soybean canopy, deep learning, improved YOLOv5s, leaf recognition, object detection

摘要： 为开发自走式大豆表型信息采集平台，实现自动化叶片识别与表型参数获取，解决人工干预问题，提出一种基于改进YOLOv5s的大豆冠层叶片识别方法。在骨干网络特征提取层中引入通道空间融合注意力机制（CBAM），提高对小叶片的识别效果。将网络中的空间金字塔池化结构（SPPF）替换为空洞空间卷积池化金字塔结构（ASPP），加强局部信息检测，引入SIoU损失函数，提高复杂背景下叶片的定位能力。利用检测框的坐标信息对单个叶片进行分割进而获取叶长、叶宽、叶面积。试验结果表明：相比于原模型，改进后模型的平均精度均值mAP、精确率、召回率分别提高5.4%、3.3%、4.6%；叶长、叶宽、叶面积平均绝对误差分别为0.98cm、0.56cm、6.20cm2。

关键词: 大豆冠层, 深度学习, 改进YOLOv5s, 叶片识别, 目标检测