Perceptual Texture Similarity Metrics.

摘要

With the ubiquity of computers and "smart" devices, it is important to obtain intuitive visual interactions between humans and computers, which include the ability to make human-like judgments of image similarity. However, pixel-based image comparisons are not suited for this task, especially when comparing texture images, which can have significant point-by-point differences, while to humans they appear to be identical. Instead, "Structural Similarity Metrics" attempt to incorporate "structural" information in image comparisons by relying on local image statistics. We develop new "Structural Texture Similarity Metrics" that are based on an understanding of human visual perception and incorporate a broad range of texture region statistics. We develop separate metrics for the grayscale component of texture and its color composition, which are attributes associated with different perceptual dimensions.;A major contribution of this thesis is a new methodology for systematic performance evaluation of texture similarity metrics, which should be targeted to each specific application. The proposed methodology considerably simplifies the testing procedures, and dramatically increases the chances of obtaining consistent subjective results. It is based on the realization that quantifying similarity when textures are dissimilar is difficult to achieve by humans or machines, and should be limited to the high end of the similarity scale. Thus, in content-based image retrieval (CBIR), the focus should be on distinguishing between similar and dissimilar textures, while in image compression, quantifying similarity should be limited to the high end of the similarity scale. For CBIR, we develop "Visual Similarity by Progressive Grouping (ViSiProG)," a new experimental procedure for subjective grouping of similar textures to serve as a benchmark for the development and testing texture similarity metrics. For image compression, we develop algorithms for generating texture deformations that facilitate subjective and objective tests at the high end of texture similarity. We also examine "known-item search" (retrieval of "identical" textures), where the ground truth can be obtained without extensive subjective tests.;Experimental results demonstrate that texture retrieval and compression performance evaluation based on the proposed metrics substantially outperforms those based on existing metrics.