Structure Similarity Metrics for Texture Analysis and Retrieval A DISSERTATION SUBMITTED TO THE GRADUATE SCHOOL IN PARTIAL FULFILLMENT OF THE REQUIREMENTS for the degree MASTER OF SCIENCE

The presence of computers in almost every sphere of today’s human life creates the need for bridging the gap between human and machine perception of the world. It is a difficult task, given that the complexity and structure of human mind and the architecture of computers are not in accordance. Since humans are mainly visual creatures, developing ways to compare the similarity of images which would mimic human perception is an important issue. It has been shown that the traditional metrics that are intrinsic to the computer representation of images, such as point-by-point comparisons, do not perform well when compared to human judgments of similarity. Therefore, there have been several attempts to incorporate structural information in image comparisons, such as Structural Similarity Metrics (SSIM). More recently, Complex-Wavelet SSIM (CWSSIM) and Structural Texture SIM (ST-SIM) have been explored in this context. The drawbacks of these approaches are that they only use the gray-scale aspect of images, and do not utilize the (usually strong) correlations between information contained in different subband decompositions of an image. In this work, an attempt is made to overcome these drawbacks and juxtapose the results to human similarity judgments for the purpose of evaluation and comparison. The information about the color composition of images is incorporated into the similarity metrics. Also, measures that represent inter-subband correlations are added to provide results closer to human evaluations.