Rice pest control effect is a key factor of rice harvest, and the prediction of pest development trend is a necessary step of pest control. To improve rice pest control effect and rice quality, a deep learning-based SD-Mask R-CNN convolutional neural network is proposed to accomplish accurate identification of segmented pest pictures. The training and testing samples of the network are collected in a multi-point manner to collect pest samples from different regions and various developmental ages to realize the diversity of rice pest samples. Based on the Mask R-CNN recognition network, and the Resnet swish DISN feature extraction network is proposed to solve the problems of gradient explosion and network degradation that easily occur in the recognition network. It is found that the recognition average accuracy AP of this recognition network reaches 98.08% and the recall rate Recall is 95.19%, which are improved by 9.8% and 17.56% respectively compared with Mask R-CNN recognition network.