Degree Name

Master of Computer Science


School of Information Technology and Computer Science - Informatics


Content-based retrieval is based on the premise that the similarity measures in the feature space accord well with visual perceptual similarity. Furthermore, the queryby-example paradigm assumes that the query concept is well specified by the user via the example image supplied. The inadequacy of these assumptions has led to the development of several similarity measures and visual features that capture and describe colour, texture and edge information in images. The simultaneous use of multiple features, relevance feedback and more recently and the use of multiple example images in specifying the query are attempts to improve the accuracy at which the query concept can be captured. Results obtained so far are still far from the ideal because of inadequate knowledge of the human perceptual processes and this leads to the so called ”Semantic Gap”. This thesis proposes a multi-image query-by-example content-based image retrieval scheme in which the significance of the components of feature vectors (intra-level) and the significance of the selected features (inter-level) are estimated through weight computation. These weights are used in calculating the feature distances and visual similarity between the query images and the database images. The hypothesis is that by incorporating the significance of features at both levels, the weighted visual similarity measure will yield improved retrieval performance (precision and recall rates). The model of the weight estimation and assignment is developed and experiments are conducted to validate the hypothesis. On average the proposed method improved the precision and recall rates in retrieval tasks on a database of natural images.

02Whole.pdf (9826 kB)