Machine Vision in Food Recognition: Attempts to Enhance CBVIR Tools

Visual identification of complex images (e.g. images of food) remains a challenging problem. In particular, contentbased visual information retrieval (CBVIR) methods, which seem a natural choice for such tasks, are often constrained by specific characteristics of the images of interest and (possibly) other practical requirements. In this paper, a novel CBVIR approach to automatic food identification is proposed, taking into account characteristics of solutions currently existing in this area. Based on limitations of those solutions, we present a scheme in which a co-occurrence of MSER features extracted from three color channels is employed to build a bag-of-words histogram. Subsequently, food images are matched by detecting similarities between those histograms. Preliminary tests on a recently published benchmark dataset UNICT-FD889 reveal certain advantages of the scheme and highlight its limitations. In particular, a need of a novel methodology for segmentation of food images has been identified.