Κατάτμηση εικόνων και εικονοσειρών για αναζήτηση βασισμένη στο περιεχόμενο

Abstract

Objective of this work is the segmentation of images and image sequences for the application of content-based retrieval and the use of the results of segmentation in new, sophisticated retrieval techniques. More specifically, this work examines the segmentation of still color images using a homogeneity-based approach, as well as the improvement of the time-efficiency of segmentation by means of the initial estimation of a coarse-grained result and its subsequent refinement. The spatiotemporal segmentation of image sequences, both uncompressed and compressed according to the MPEG-2 standard, is also examined, emphasizing in both cases on the efficient use of motion information. The resulting image and image sequence segmentation algorithms are subsequently utilized in the development of a new retrieval technique which combines them with a simple ontology and a relevance feedback mechanism, avoiding the adoption of restrictive assumptions and leading to high retrieval performance. Finally, linking the need for comparative evaluation of segmentation algorithms with the application of retrieval, a known retrieval paradigm and a set of standardized descriptors are used for evaluating the effect of the developed segmentation algorithms on retrieval performance.