It was difficult to take high-quality images when pests were still and in close distance in the natural environment of rice field, which led to the problem that satisfactory identification accuracy could not be achieved when using the actual environmental identification model detection. A low-resolution rice pest image recognition method based on self-calibrated convolutions and ResNeSt block for ResNet50 (SCResNeSt) was proposed. Firstly, the enhanced super-resolution generative adversarial networks (ESRGAN) super partition network was used to enhance the data of low-resolution images to solve the problem of less effective information about rice pests. In SCResNeSt network, three consecutive 3×3 convolutional layers were used to replace the first 7×7 convolutional layer to reduce the computational cost. Using self-calibrated convolution instead of the 3×3 convolution in layer 2, through internal communication, the field of view of each convolutional layer was explicitly extended to obtain part of the background information of pest images, to enrich the output features. The split-attention network block (ResNeSt block) was used in the backbone network to further improve the accuracy of obtaining pest information in the image. Finally, the optimized model was deployed on the mobile terminal, and a lightweight mobile rice pest identification system was developed. The experimental results showed that compared with the existing methods, the ESRGAN model could recover the real information about crop pests, and the SCResNeSt model could effectively improve the performance of rice pest identification, the accuracy can reach 91.20%, which showed that the depth model could accurately identify rice pest types. The research result can provide an important technical basis for the intelligent identification and control of rice pests, and it would improve the level of rice production informatization.