An improved statistical texture residual learning network （STRLNet） for the fine-grained classification of grape leaves was constructed to solve the problem of low classification accuracy of intra-class varieties caused by high inter-class similarity between grape leaves. SE attention mechanism was added on the basis of ResNet50 backbone network. The feature enhancement layer of the underlying information was built. The enhanced underlying features with the high-level semantic information extracted from the backbone network were integrated. The output was connected to the full connection layer used for storing the characteristics of classification. The collected dataset of mature grape leaves of 11 cultivars were used for training and testing. The results showed that STRLNet fully utilized the underlying feature information while improving the spatial performance of the network， with a classification accuracy of 92.26% for the collected dataset of grape leaves. It was about 2.8 percentage points higher than that of the ResNet backbone network. It had higher accuracy in fine-grained classification of grape leaves compared with mainstream classification networks including VGG16， Inception v4， and ResNet. It is indicated that the improved model can focus on more feature information compared with the backbone network in the classification of grape leaves from multi-cultivars. It can achieve higher classification accuracy compared with the mainstream classification network models and further improve model performance.