Compressed domain object segmentation and indexing.

摘要

In this thesis, compressed-domain object segmentation and indexing algorithms are proposed to achieve fast segmentation and indexing of objects of interest. Compressed-domain information is used as primary input to segment and to generate the indices of the objects. By using compressed data without inverse transformation, the amount of information needed to be processed and the complexity of the segmentation and indexing algorithms can be greatly reduced. To efficiently segment objects of interest, we propose two novel approaches in segmenting objects directly in compressed domain. These are the modified region merging and the adaptive-threshold region merging. By using the modified region merging technique, the regions are gradually merged from high similarity to low similarity among their neighbors. This technique gives better segmentation results than using the other technique. However, selecting the optimum thresholds for each video sequence is a very difficult task since each video sequence has different characteristics. Even though the segmentation results obtained from the adaptive-threshold region merging may contain errors at the boundary, the implementation is fast and simple. After the segmentation process, several features can be extracted from the segmented object. The shape and color features are adopted in the proposed object indexing system. Three shape matching techniques B-spline several features can be extracted from the segmented object. The shape and color features are adopted in the proposed object indexing system. Three shape matching techniques B-spline knot matching, matching using Fourier descriptors, and matching using invariant Fourier descriptors, are investigated. Shape matching using invariant Fourier descriptors is selected as the shape matching technique in the proposed indexing system because of its robustness and invariant properties to rotation, scale, translation, affine transform, and mirror effect. Since most of video sequences in the database contain human objects, we employ face shape and the object contour as our combined shape features to separate human from non-human objects. Using the combined shape features, the retrieval results are greatly improved. The color features are incorporated with the shape features to further improve the proposed indexing system in case of different objects having similar shapes. By using these proposed combined shape and color features, the retrieval performance is improved significantly. Hence, these combined features are appropriate for indexing of objects of interest and are employed in the proposed indexing system.