基于轻量化GMNC-YOLO的苹果叶片病害检测方法

Apple leaf disease detection method based on lightweight GMNC-YOLO

针对苹果叶片病害识别中受人为主观因素影响大、传统卷积神经网络识别精度低、网络模型参数量过大等问题,文章提出了一种轻量化GMNC-YOLO检测算法对苹果叶片病害进行检测。该算法以YOLOv5s为基础模型,使用Ghost卷积进行特征提取,以实现网络轻量化;采用MobileOne模块替换Neck的C3模块,结合NAM与CBAM设计新的注意力机制模块NCB,以充分挖掘叶片图像中的信息,并提高模型检测精度。实验结果表明,相较于YOLOv5s,改进后的苹果叶片病害检测方法的平均精度提高了3.0%,参数量、FLOPs及权重文件大小分别降低约29.8%、32.5%和26.8%;与当前主流算法对比,GMNC-YOLO具有一定的先进性。

In order to solve the problems of strong subjective factors,low recognition accuracy of traditional convolutional neural network and too large parameters of network model in apple leaf disease recognition,a lightweight GMNC-YOLO detection algorithm was proposed to detect apple leaf disease.The algorithm is based on YOLOv5s model,and features are extracted by Ghost convolution to realize network lightweight.MobileOne module is used to replace C3 module of Neck,and a new attention mechanism module NCB is designed by combining NAM and CBAM,which can fully mine the information in leaf images and improve the accuracy of model detection,The results show that compared with YOLOv5s,the average accuracy of the improved apple leaf disease detection method is improved by 3.0%,and the parameters,FLOPs and weight file size are reduced by about 29.8%,32.5%and 26.8%respectively,and GMNC-YOLO has a certain advanced compared with the current mainstream algorithms.