A Novel Method for Exploring Patch-level Context to Improve Image Categorization Performance

概要

論文の詳細を見る
Patch-level context can be used to relieve ambiguity of patch encoding in the bag of visual words method. However, in previous research patches of interest are often combined to their context in a fixed way. In this paper, we propose a novel approach in which the variation of the closeness of the relationship between a patch of interest and its context is taken into consideration, so that the context can be utilized more effectively. Specifically, we associate each patch of interest with other patches and take them as its context. In patch encoding, the patch of interest is first encoded by its N nearest visual words. Then, the encoding is adjusted based on its context by investigating whether its assignments to the visual words get support from the context patches. Furthermore, to explore the image content extensively, given an input image, we create three representations using different context strategies and fuse them in a probabilistic manner. Finally, we evaluated the proposed method on the datasets scene categories 8, scene categories 13 and scene categories 15, respectively. Experimental results demonstrated the effectiveness of the proposed method.
電気学会の論文