Learn More for Food Recognition via Progressive Self-Distillation

Food recognition has a wide range of applications, such as health-aware recommendation and self-service restaurants. Most previous methods food firstly locate informative regions in some weakly-supervised manners then aggregate their features. However, location errors limit the effectiveness these to extent. Instead locating multiple regions, we propose Progressive Self-Distillation (PSD) method, which progressively enhances ability network mine more details for recognition. The training PSD simultaneously contains self-distillations, teacher student share same embedding network. Since receives modified image from its by masking outputs stronger semantic representations than Guided with semantics, is encouraged useful enhancing own ability. also enhanced shared By using progressive training, incrementally improves discriminative regions. In inference phase, only used without help Extensive experiments on three datasets demonstrate our proposed method state-of-the-art performance.