Texture Representations for Content-based Retrieval

M.S. Thesis Kun Xu


Content-based image retrieval systems are of great interest today because of the need for efficient ways to search large databases of digital images. These systems employ low-level features such as texture, color and shape to characterize the salient information in the image, and the matching between the images is evaluated by computing similarity measures. Using texture as an example, we have investigated the quality of representation provided by low-level features for the semantic meaning of the image. We have shown that a more accurate description of the underlying distribution of low-level features does not improve the retrieval performance. We have also introduced a simplified multiresolution symmetric autoregressive model to describe textures, and a similarity measure based on the Bhattacharyya distance. The effectiveness of the new autoregressive model and distance measure was compared with that of the texture representations employed in the literature: Wold, multiresolution simultaneous autoregressive (MRSAR) model and Gabor filter bank, and the Mahalanobis distance based similarity measure. Experiments were performed on all the available texture databases: Brodatz, VisTex and MeasTex. The issue of homogeneity of an image was examined by defining a database containing the 50 perceptually most uniform images from the Brodatz database. To facilitate the experimental work a prototype texture retrieval system was developed which allows the user to search by using different combinations of texture representations and distance measures.

The thesis has part1 and part2. The size of the compressed files is about 12 M. The thesis contains 60 pages.

Return to Theses