Color image retrieval is to search color images using queries represented by image descriptors, which usually describe color distribution and relation of color pixels in an image. A color co-occurrence histogram (CCH) among the descriptors captures information on the spatial layout of colors within an image. It has shown excellent performance on color image retrieval, but requires many bins to describe contents of images and has bad effect on the similarity of same contents images, in which the size of homogeneous color regions are highly different.
To resolve these problems and to improve retrieval performance, this thesis proposes a weighted CCH and two image retrieval methods using it. Generally the process of image retrieval using a CCH has three steps. The first step is to get the CCH from a query image. The second step is to compute similarity between CCHs of the query image and reference images. The last step is to sort reference images by the similarities and to visualize them. The proposed retrieval methods weight main diagonal and off-diagonal elements of a CCH in the first and/or the second steps mentioned above.
Experiments have shown that the proposed methods with a few bins outperform some conventional methods when large weight is given on off-diagonal elements regardless of color quantization levels. We believe that the effectiveness of the method is caused by the characteristics describing the size and the coherence of homogeneous color regions and being robust to size variation of the color regions. Moreover, the image retrieval performance is little affected by the threshold, which is an energy level of valid bins, regardless of color quantization levels.
The proposed methods use contents of images effectively, so they can be effectually used in the other content-based applications such as color image classification, color object tracking, and video cut detection.