Foodness Proposal for Multiple Food Detection by Training of Single Food Images

Abstract

We propose a CNN-based "food-ness" proposal method which requires neither pixel-wise annotation nor bounding box annotation. Some proposal methods have been proposed to detect regions with high "object-ness" so far. However, many of them generated a large number of candidates to raise the recall rate. Considering the recent advent of the deeper CNN, these methods to generate a large number of proposals have difficulty in processing time for practical use. Meanwhile, a fully convolutional network (FCN) was proposed the network of which localizes target objects directly. FCN saves computational cost, although FCN is essentially equivalent to the sliding window search. This approach made large progress and achieved significant success in various tasks. Then, in this paper we propose an intermediate approach between the traditional proposal approach and the fully convolutional approach. Especially we propose a novel proposal method which generates high "food-ness" regions by fully convolutional networks and back-propagation based approach with training food images gathered from the Web.