Food Image Recognition Method Based on Generative Self-supervised Learning.

The demand of social life for automatic recognition of food images is increasing. Food images have the characteristics of diverse forms, small differences between classes and large differences within classes, which has the problem of high recognition difficulty. This paper proposes a food image recognition method based on generative self-supervised learning. Firstly, we use a BEiT based pre-training model which is trained through generative self-monitoring learning method as the feature extraction network to extract the global semantics and local detail features of food images. And then we fine-tune the fully connected network MLP for classification and recognition through supervised learning method. The model is tested on the current mainstream public food image dataset Food-101, and the top-1 accuracy of 85.99% is obtained. The experimental results show that this method can significantly reduce the computation of pixel level expression as well as extract the global and detailed features of the image, achieving quite good food image classification and recognition effect. Our method has good robustness, generalization and flexibility, which has practical application value.