Superpixel-based Image Recognition for Food Images

Food is an inseparable part of people's lives. Food image recognition has been attracting increasing attention due to the advances of Internet, imaging techniques and social media. Approaches for food image recognition are mainly focused on two main directions: low-level approaches and mid-level approaches. Low-level approaches extract low-level local features, such as SIFT or SURF, following feature encoding techniques. Mid-level approaches extract higher-level image parts and have shown promising results in many recognition problems. Compared with other image recognition problems, food images are highly deformable with large infra-class variance and small between-class variance. In this paper, considering that mid-level approaches' superior performance and superpixels segmentation methods' ability to successfully segment food parts, we propose a superpixel-based food image recognition framework to mine mid-level superpixel food parts-to-class similarity. We evaluate the proposed framework on UEC Food dataset, and show promising results when compared with existing state-of-the-art methods.