DietCam: Regular Shape Food Recognition with a Camera Phone

The purpose of this paper is to develop an automatic camera phone based multi-view food classifier as part of a food intake assessment system. Food intake assessment is important for obesity management, which has shown significant impacts in public healthcare. Conventional dietary record based food intake assessment methods exhibit insufficient popularity due to their low accuracy and high dependence on human interactions. Image based food recognition appears recently. But it is still under development and far away from field applications. This paper presents DietCam, a camera phone based application to evaluate food intakes automatically from multiple perspectives. Food recognition from images is afflicted currently with a low recognition accuracy caused by the uncertainties of food appearances. The deformable nature of food items together with the complex background environment makes the problem even harder. DietCam separates every food item through evaluating the best perspective and recognize each of them from multiple images with a probabilistic method. The recognition accuracy is increased through an enhanced joint distribution from every viewpoint. A prototype of DietCam has been implemented on iPhone. In the field experiments, it shows an accuracy of 84% for regular shape food items.