Abstract: A high-performance real-time fine-grained plant disease detection framework is proposed to solve the problems of dense distribution, irregular shape, multi-scale target category and texture similarity encountered by existing identification methods in plant disease detection. Firstly, two new residual blocks are designed in YOLOv8 backbone network and neck to enhance feature extraction and reduce computing cost. Secondly, the DenseNet layer is introduced and the Hard-Swish function is used as the main activation function to improve the accuracy of the model. Finally, the PANet network is designed to retain finegrained local information and improve feature fusion. Four different diseases of tomato plants were detected in different complex environments, and the experimental results showed that the proposed model was superior to the most advanced detection models in both accuracy and speed. At the detection rate of 71.23 FPS, the model obtained the precision of 92.58%, the recall rate of 97.59%, and the F1score of 93.64%, which provided an effective technical means for precision agriculture automation.

Key words: plant disease detection, improved YOLOv8, realtime target detection, deep neural network, residual network

摘要： 为解决现有识别方法在植物病害检测中遇到的密集分布、不规则形态、多尺度目标类别、纹理相似性等障碍，提出一种高性能的实时细粒度植物病害检测框架。首先，在YOLOv8主干网络和颈部设计两个新的残差块，增强特征提取和降低计算成本；其次，引入DenseNet层，并使用HardSwish函数作为主要激活函数，以提高模型的准确性；最后，设计PANet网络，用于保留细粒度的局部信息和改善特征融合。在不同的复杂环境下，对番茄植株的四种不同病害进行检测。试验结果表明，所提改进模型在检测准确性和速度上均优于现有模型的检测模型。当检测速度为71.23 FPS时，所提改进模型精确度为92.58%，召回率为97.59%，F1分数为93.64%。为精准农业自动化提供有效的技术手段。

关键词: 植物病害检测, 改进YOLOv8, 实时目标检测, 深度神经网络, 残差网络