MFNet: A Channel Segmentation-Based Hierarchical Network for Multi-food Recognition

Due to significant differences between foods from different regions, there is a relative lack of publicly available multi-food image datasets, which is particularly evident in the field of Chinese cuisine. To alleviate this issue, we create a large-scale Chinese food image dataset, JNU FoodNet, which contains 17,128 original images. Although a considerable amount of prior research has focused on single food image recognition, such methods are not suitable for recognizing multiple food items in one image. Moreover, in a food image, there are usually multiple food regions, and the key features of each region tend to gradually disperse from the center to the edges, leading to cumulative errors. To overcome this difficulty, we design a selective discriminative feature constrained module, SGC, which restricts model attention to regions from a global information perspective. Furthermore, we propose a progressive hierarchical network, MFNet, based on channel segmentation from both the whole image and local region perspectives, combined with the SGC branch. Experimental results show that MFNet achieves state-of-the-art mAP values on JNU FoodNet, UEC Food-100, and UEC Food-256.