CNN-based visible ingredient segmentation in food images for food ingredient recognition

Food ingredient identification aims to predict the ingredient composition of the food image. It is a challenging task due to the large intra-class variance and small inter-class variance. In this work, we propose a standard hierarchical ingredient structure and base on this structure, and introduce a hierarchical multi-label single-ingredient dataset. We further propose a hierarchical and a non-hierarchical ingredient segmentation framework to partition ingredients from multi-ingredient food images for ingredient identification. Furthermore, we introduce three types of backbones for the ingredient segmentation. Our experiments are conducted on the hierarchical multiple ingredient test dataset for each level. Experimental results verify the possibility of the food ingredient segmentation with weakly supervised learning for further food ingredient identification.