事前知識に基づく食事領域抽出の改良と実画像を用いた比較評価

抄録

In this paper, we propose a method which can extract food regions from food images to improve extraction accuracy compared with a conventional method by decreasing misdetermination of the food regions as the background regions. The proposed method uses a convex hull based on the local extrema and their density to generate initial seeds for GrabCut, which can revise the food and background regions on the basis of color similarity and distribution. Our experiment demonstrated that the proposed method significantly increased the F-measure, which shows the comprehensive extraction accuracy, by 4.41% or more compared with the conventional method. The proposed method also increased the F-measure by 4.54% or more compared with SegNet based on deep convolutional neural network trained by 1017 food images available on the Internet. These results provided the fact that the proposed method was effective in the extraction accuracy compared with the existing methods which can be constructed by limited introduction cost.