Category pattern mining based image retrieval

Abstract

Image retrieval is to find out the similar semantic images to the query image, which is an important task in the field of image recognition. It is still an open challenging task due to the semantic gap of image understanding. The traditional image retrieval method is a simple retrieval between the query image and the database. However, only a query image contains weaker category information, so that the traditional image-based retrieval results are not satisfactory. In this paper, we propose a category pattern mining (CPM) strategy to extend an image (point) to an image category (plane). It means the semantic extension is performed from the individual query image to the whole image category. The proposed PTP (point to plane) method mined the category pattern of the query image and enriched the semantic information. The main contribution of the PTP framework is to improve the image retrieval from the traditional image-based retrieval into the new category-based retrieval. Experimental results and evaluations on two databases demonstrate that the proposed PTP method achieves an obvious superiority in the image retrieval tasks.