Objective: Improve the speed and accuracy of foreign matter identification in food. Methods: Based on the LeNet-5 network structure, the improved CNN model was obtained by adding batch normalization layer and dropout layer. Using this model, a recognition system was established for the automatic recognition of foreign bodies in food images. The performance of the model was analyzed through experiments. Results: Compared with the traditional model, this model has higher detection accuracy and faster recognition speed. The recognition accuracy of food foreign bodies was 99.75% and the recognition time was only 0.332 s. Conclusion: The foreign object recognition model of dumpling image had good detection speed and recognition accuracy.