关键词: 番茄, 病害, 图像识别, 卷积神经网络, 深度残差网络

Abstract: Aiming at the problems of large memory consumption and large computational resource consumption in most Deep Convolutional Neural Network（DCNN）models，a recognition method of tomato disease images based on improved Deep Residual Network（DRN）was proposed. Based on the traditional residual neural network，the proposed network model used multi-scale convolutions to replace the single-scale convolutions in the original network architecture，which made the extracted features richer，expanded the network width and avoided the degradation problem caused by excessively deep network. In order to further reduce the memory requirement of the model，partial standard convolutions were replaced by depthwise separable convolutions，and the model parameters were reduced without degrading network performance. To verify the effectiveness of the improved deep residual network model for tomato disease recognition，the limited dataset of tomato disease leaf images was augmented. Based on the augmented dataset，the improved model was compared with several commonly used deep neural network models. The results show that the improved DRN model can identify tomato diseases well，which achieves a mean accuracy of 98. 58% on test dataset，and has the size of only 19. 0 MB. So the proposed model will help to achieve real-time diagnosis of tomato diseases on low-performance terminals in the future.

Key words: tomato, disease, image recognition, Convolutional Neural Network (CNN), Deep Residual Network (DRN)