Textural Features For Content-Based Image Database Retrieval

Program Authorized to Offer Degree Date In presenting this thesis in partial fulfillment of the requirements for a Master's degree at the University of Washington, I agree that the Library shall make its copies freely available for inspection. I further agree that extensive copying of this thesis is allowable only for scholarly purposes, consistent with " fair use " as prescribed in the U.S. Copyright Law. Any other reproduction for any purposes or by any means shall not be allowed without my written permission. Image database retrieval has received significant attention in recent years. This thesis describes a system to retrieve all database images having some section similar to the query image. We develop efficient features for image representation and effective metrics that use these representations to establish similarity between images. The first set of features we use are the line-angle-ratio statistics constituted by 2-D texture histograms of the angles between intersecting/near-intersecting lines and the ratios of mean gray levels inside and outside the regions spanned by those angles. A line selection algorithm using hypothesis testing is developed to eliminate insignificant lines. The second set of features used are the variances of gray level spatial dependencies computed from co-occurrence matrices at different distances and ori-entations. Statistical feature selection methods are used to select the parameters of the feature extraction algorithms. We also combine these macro and micro texture features to make use of their different advantages. We define two classes, the relevance class and the irrelevance class, and design an automatic groundtruth construction protocol to associate each image pair with one of the classes. To rank-order the database images according to their similarities to the query image, a likelihood ratio and the k-nearest neighbor rule are used. To evaluate the performance, classification effectiveness and retrieval performance experiments are done on a large database with many different kinds of complex images. More than 450,000 image pair classifications using a Gaussian classifier and a nearest neighbor classifier showed that approximately 80% of the relevance class groundtruth pairs were assigned to the relevance class correctly. To compensate for the effects of the mislabeling probabilities of the groundtruth construction protocol, we develop a statistical framework that estimates the correct classification results. Hence, some of the assignments which we count as incorrect are not in fact incorrect. In the retrieval performance tests, which use more than 300,000 queries, we observed that combined feature sets and …