SSDB-Net: A Single-Step Dual Branch Network for Weakly Supervised Semantic Segmentation of Food Images

Qingdong Cai, Charith Abhayaratne

Abstract: Food image segmentation, as a critical task in food and nutrition research, promotes the development of various application domains such as calorie and nutrition estimation, food recommender systems, and daily food monitoring systems. Currently, most of the research is focused on food and non-food segmentation, which simply segments the food and background regions. Differently, semantic food segmentation can identify different specific food ingredients in a food image and provide more detailed and accurate information such as object location, shape and class. This is a more challenging but meaningful task, because the same food may appear in completely different colours, shapes and textures in different dishes, and correspondingly less researched. From the implementation perspective, most previous research is based on deep learning methods with pixel-level labelled data. However, annotating pixel-level labels requires extremely high labour costs. In this paper, a novel single-step dual branch network (SSDB-Net) is proposed to achieve weakly supervised semantic food segmentation. To our knowledge, this research is the first time proposing weakly supervised semantic food segmentation with image-level labels based on convolutional neural networks (CNN). It may serve as a benchmark for future food segmentation research. Our proposal method resulted in an mIoU of 14.79% for 104 categories in the FoodSeg103 dataset compared to 11.49% of the state-of-the-art WSSS method applied in food domains.