Image segmentation of persimmon leaf diseases based on UNet

Abstract

The size and shape of lesion area of persimmon diseases will vary with the occurrence period and degree. CNN has fixed receptive field when extracting persimmon disease features, which can not adapt to the geometric changes of disease spots, resulting in incomplete disease feature extraction and reducing the segmentation accuracy of persimmon disease images. Deformable convolution can dynamically adjust the size of the receptive field according to the input features, and automatically adapt to the geometric deformation of the lesion. This paper proposes a UNet based on self-attention mechanism and deformable convolution for image segmentation of persimmon leaf disease. With UNet as the basic network, the standard convolution in the down-sampling stage of UNet is replaced by deformable convolution to extract more abundant features, and the self-attention mechanism is used to learn the relationship between the various features to obtain more spatial information and context information. The experimental results show that the mPA and the mIoU of the proposed algorithm are 89.18 % and 83.58 %, its segmentation effect is better than UNet.