Image Segmentation Model of Pear Leaf Diseases Based on Mask R-CNN

Image segmentation of pear leaf diseases is a crucial method to scientifically determine the occurrence degree of diseases. Pear leaf diseases sizes differ a lot affected by the occurrence period and degree. Convolutional neural network has a fixed receptive field when extracting pear leaf diseases images, which makes it difficult to fully extract diseases features of different sizes. At the same time, continuous down-sampling reduces the resolution of the feature map and loses a lot of spatial information, resulting in rough mask of the image restored by upsampling. To solve this problem, a pear leaf diseases image segmentation model based on Mask R-CNN is proposed. In the feature extraction stage, the dilated convolution is used to control the expansion rate of the convolution kernel to gain different sizes of the receptive field. Under the condition of retaining the resolution of the feature map, multi-scale context information is obtained to improve the accuracy of mask prediction. The experimental results show that the MPA and MIoU of the proposed algorithm are 91.49 % and 88.55 % respectively, and the segmentation effect is better than that of Mask R-CNN.