A Structured Committee for Food Recognition

Abstract

Food recognition is an emerging computer vision topic. The problem is characterized by the absence of rigid structure of the food and by the large intra-class variations. Existing approaches tackle the problem by designing ad-hoc feature representations based on a priori knowledge of the problem. Differently from these, we propose a committee-based recognition system that chooses the optimal features out of the existing plethora of available ones (e.g., color, texture, etc.). Each committee member is an Extreme Learning Machine trained to classify food plates on the basis of a single feature type. Single member classifications are then considered by a structural Support Vector Machine to produce the final ranking of possible matches. This is achieved by filtering out the irrelevant features/classifiers, thus considering only the relevant ones. Experimental results show that the proposed system outperforms state-of-the-art works on the most used three publicly available benchmark datasets.