Abstract: Early detection of tomato leaf disease category is conducive to rapid diagnosis and treatment to save crop losses.The traditional method of tomato disease identification based on deep learning has the problems of large model size and large consumption of computing resources, which is not suitable for direct deployment on portable devices with low computing power and limited storage space.In this study, knowledge distillation technology was used to compress the model, and pyramid squeeze attention module was used to improve the teacher network ResNet50 to enhance the network performance.Under the guidance of the teacher network, the student network ShuffleNetV2 had achieved excellent performance.By selecting tomato diseased leaves in PlantVillage dataset for experiment, the experimental results showed that the distilled network KD-ShuffleNetV2 improved the accuracy of the model, and saved more storage space and computing resources than deep convolution neural networks Alexnet, Vgg11,and ResNet50.The network achieved 95.66% recognition accuracy on tomato disease dataset, and the size of the model was only 4.98 MB.Finally, the model was transplanted and deployed to the low-cost Raspberry Pi to complete the tomato leaf recognition system development and recognition application.