Partition-based Color Image Retrieval

This paper provides a new image retrieval approach: Partition-based color image retrieval. Partition is used to extract spatial information and by coding, image can be transformed into a text-image. So images can be analyzed by mature text model.Color is a common used feature in content-based image retrieval for it simplifies object identification. The traditional approach to using the color information of an image is color histogram, which is insensitive to translation and rotation. But its drawback is prone to yield false hits in large database because of lacking spatial information. To solute the problem, we try to combine spatial information with color feature to improve the performance of content-based image retrieval. This is achieved by partitioning images in the training set into fixed size cells and, for each cell, extracting a local color histogram as the color invariant feature of the cell. All of the color invariant features are clustered into a number of patterns. Thus all the images in the database can be regarded as a collection of those patterns. Images are recognized by their patterns, which takes a step to retrieve image by symbolic notions. Thereby the well-developed text retrieval method can be applied for image'query and index through such symbolic descriptions. Experimental results show that the new method is robust in retrieving images with domain-free scenes and is efficient in sub-region retrieval and localization.For all the partition-based image retrieval, choosing a proper partitioning scale is a significant problem and is usually determined subjectively. This paper proposes information entropy as a measure of optimal scale and experimental results show it is reasonable.