农业机械学报2024，Vol.55Issue(4)：204-212,9.DOI:10.6041/j.issn.1000-1298.2024.04.020
基于CycleGAN-IA方法和M-ConvNext网络的苹果叶片病害图像识别
Image Recognition of Apple Leaf Disease Based on CycleGAN-IA Method and M-ConvNext Network
摘要
Abstract
Aiming at the problems of difficult dataset acquisition,insufficient samples,and low recognition accuracy in apple leaf disease image recognition,a disease recognition network based on multi-scale feature extraction ConvNext(M-ConvNext)model was proposed.A data enhancement method combining improved CycleGAN and affine transformation(CycleGAN-IA)was used.Firstly,the CycleGAN network structure was optimized by using a convolutional kernel with a smaller sensory field and a residual attention module,and a binary cross-entropy loss function instead of the mean-variance loss function of CycleGAN network,in order to generate high-quality sample images and improve the complexity of sample features;then affine transformation was applied to the generated images to improve the spatial complexity of the data samples,which solved the problem of insufficient data samples,and was used to assist the subsequent disease recognition model.Secondly,the M-ConvNext network was constructed,which was designed with the G-RFB module to acquire and fuse the feature information of each scale,and the GELU activation function enhanced the feature expression ability of the network to improve the accuracy of apple leaf disease image recognition.Finally,the experimental results showed that the CycleGAN-IA data enhancement method can play a good role in expanding the dataset,and it was verified on the commonly used network that the enhanced dataset can effectively improve the accuracy of apple leaf disease image recognition;through the ablation and comparison experiments,the recognition accuracy of M-ConvNex can be up to 99.18％,which was 0.41 percentage points more than the original ConvNext network,and 3.78 percentage points,7.35 percentage points,4.07 percentage points higher than that of ResNet50,MobileNetV3,and EfficientNetV2 networks,respectively,which provided an idea and laid a foundation for the subsequent recognition of crop diseases.关键词
苹果叶片/病害识别/生成式对抗网络/数据增强/多尺度特征提取Key words
apple leaf/disease identification/generative adversarial networks/data enhancement/multi-scale feature extraction分类
农业科技引用本文复制引用
李云红,张蕾涛,李丽敏,苏雪平,谢蓉蓉,史含驰..基于CycleGAN-IA方法和M-ConvNext网络的苹果叶片病害图像识别[J].农业机械学报,2024,55(4):204-212,9.基金项目
国家自然科学基金项目(62203344)、陕西省科技厅自然科学基础研究重点项目(2022JZ-35)和国家级大学生创新创业计划项目(202210709012) （62203344）
