Abstract: Since the initial symptoms of Peach Tree Yellow Leaf Disease (PTYLD) are not readily apparent, the existing deep learningbased recognition techniques for this disease suffer from issues like inaccurate recognition and limited recognition species. To address this, a recognition model of PTYLD based on FasterRCNN (Regionbased Convolutional Neural Network) is proposed. In order to enhance the recognition accuracy and diversity of PTYLD, RSLoss function is used to replace the crossentropy function in the Region Proposal Network (RPN), and the SoftNMS algorithm is used to replace the original NonMaximum Suppression (NMS) algorithm, so as to improve FasterRCNN. The recognition effect  of the initial and improved version of FasterRCNN models on PTYLD is compared by experiments. The experimental results demonstrate that the improved FasterRCNN achieves a mean average precision (mAP) of 90.56%, recall rate of 94.16%, an accuracy of 92.53% for each category of yellow leaf disease,  and can identify five common PTYLD.

Key words: peach tree yellow leaf disease, FasterRCNN, complex background, SoftNMS

摘要： 由于桃树黄叶病（以下简称PTYLD）初期症状不明显，现有的基于深度学习的桃树病害识别技术，存在识别准确率不高、识别品种单一的问题，提出一种基于FasterRCNN的PTYLD识别模型。为提高模型对PTYLD识别准确率和识别多样性，提出使用RSLoss函数代替RPN中的交叉熵函数、使用SoftNMS算法代替原来的NMS算法，来改进FasterRCNN。通过试验对比初始版和改进版FasterRCNN对PTYLD的识别效果。试验结果显示，改进后的FasterRCNN对黄叶病识别的各类别平均准确率mAP达90.56%、召回率达94.16%、准确率达92.53%，能识别常见的五种PTYLD。

关键词: 桃树黄叶病, FasterRCNN, 复杂背景, 软性非极大值抑制算法