Chromatic invariant based image retrieval for three dimensional objects

Abstract

This thesis presents new methods for object-based image retrieval and extraction using color, texture, and spatial information. Object-based image representation plays an important role for new generation of digital video due to the explosive growth of multimedia data. Extracting semantically meaningful objects from an image of natural scene is still a formidable task for many multimedia applications such as content-based image retrieval, digital library, and interactive television. Furthermore, new video coding standards provide the user with flexibility for object-based access and manipulation of video data, while the production of these contents is out of the scope of the standards. Thus, the success of any object-based approach depends largely on the extraction of the object regions. To solve these problems, firstly, this thesis presents an efficient indexing/matching algorithm that is independent of the changes in the illuminant color and the geometric conditions for three-dimensional objects with multiple colors. The color contents of an object can be represented by the chromaticities corresponding to the distinct color regions in an image. However, the change in illumination color would result in a deformation of the chromaticity distribution so that it degrades the performance of color recognition. In order to discount the dependency due to illumination color, we define a color constant descriptor called a chromatic invariant which is stable to the changes in illumination and geometrical conditions. For the object-based interactive broadcasting, we develop an algorithm for the region-based object extraction and image retrieval system. When an object of interest is given as a query image, this algorithm retrieves the images containing the object of interest and extracts the object regions from the retrieved images automatically. Extracting local features reliably, which is necessary for the effective representation of the local regions is generally more ...