Abstract:

Aiming at the problem of poor segmentation and recognition of apple leaf spots under natural conditions, this paper proposed a semantic segmentation model for apple leaf diseases that incorporated conditional random fields and convolutional block attention modules to achieve accurate segmentation and recognition of spots of apple leaf rust, brown spot, gray spot and Alternaria leaf spot disease. In this paper, based on the U-Net model, ResNet50 was used as the backbone network to prevent the gradient vanishing problem, and the convolutional block attention module was added to the jump-connected branch and the up-sampling layer respectively, to reduce the loss of segmentation accuracy during the training process, and the fusion of dice loss and the cross-entropy loss function to reduce the loss fluctuation, and finally, the segmentation results were optimized using the conditional random field to obtain the diseased spot mask image, which was used to realize the accurate segmentation and recognition of apple leaf rust, brown spot, grey spot and Alternaria leaf spot disease so as to realize semantic segmentation of apple leaf diseases. In this study, we conducted experiments on the homemade apple leaf disease dataset, and analyzed the effects of light, shadow and water droplets on the segmentation results. The experimental results showed that the semantic segmentation model constructed in this paper improved the average segmentation accuracy mIoU by 8.24 percentage points, the average classification accuracy mPrecision by 11 percentage points, and the average pixel accuracy of category mPA by 6.09 percentage points compared with the traditional U-Net model, and was less affected by uneven illumination and raindrops, and had better robustness and reliability.

Key words: disease segmentation, attention mechanism, conditional random field, deep semantic segmentation