In order to improve the accuracy of food classification in practical scenarios, a new food data set construction and network improvement method was proposed. Through the construction of practical dataset, data augmentation, model enhancement and model validation, a set of food image classification technology was established. In terms of model enhancement, the CBAM was added to DenseNet264 while retaining the features learned from ImageNet to improve the model's attention to key features and its generalization ability. A comparison between the standard dataset and the augmented dataset showed that the improved model achieved accuracy rates of 88.43% and 91.8%, representing a 5.28% and 4.74% increase over DenseNet264, respectively. The precision rates were 87.5% and 90.98%, the recall rates were 88.01% and 90.65%, and the F1 scores were 87.75% and 90.81%. On the augmented dataset, the accuracy of each network was improved by an average of 3.11% compared to the standard dataset, with DenseNet264 showing the best improvement. CBAM and transfer learning significantly enhanced the feature extraction capability and accuracy of the model, leading to better performance in the task of classifying practical datasets and having certain value in practical application. This technology will be applied in practical catering systems to provide a better user service.