In view of the fact that traditional neural networks cannot effectively classify Chinese food with high similarity, a Chinese food recognition model of RNA-TL (ResNet with attention and triplet loss) based on an improved residual network is proposed. The algorithm first fuses the multi-scale features to extract the semantic information of deep-level images, and then adds an attention mechanism layer to give more attention to the important parts of the images. Finally, the similarity among classes is calculated by using triplet-loss, whose result is input into support vector machine (SVM) for classification. The experimental results indicate that the proposed RNA-TL model possesses more superior performances in recognition accuracy on the public dataset of Chinese food and the dataset collected by our project team, compared with the other mainstream algorithm models.