Perceptual Image Similarity Metrics and Applications.

Abstract

This dissertation presents research in perceptual image similarity metrics and applications, e.g., content-based image retrieval, perceptual image compression, image similarity assessment and texture analysis.

The first part aims to design texture similarity metrics consistent with human perception. A new family of statistical texture similarity features, called Local Radius Index (LRI), and corresponding similarity metrics are proposed. Compared to state-of-the-art metrics in the STSIM family, LRI-based metrics achieve better texture retrieval performance with much less computation. When applied to the recently developed perceptual image coder, Matched Texture Coding (MTC), they enable similar performance while significantly accelerating encoding. Additionally, in photographic paper classification, LRI-based metrics also outperform pre-existing metrics. To fulfill the needs of texture classification and other applications, a rotation-invariant version of LRI, called Rotation-Invariant Local Radius Index (RI-LRI), is proposed. RI-LRI is also grayscale and illuminance insensitive. The corresponding similarity metric achieves texture classification accuracy comparable to state-of-the-art metrics. Moreover, its much lower dimensional feature vector requires substantially less computation and storage than other state-of-the-art texture features.

The second part of the dissertation focuses on bilevel images, which are images whose pixels are either black or white. The contributions include new objective similarity metrics intended to quantify similarity consistent with human perception, and a subjective experiment to obtain ground truth for judging the performance of objective metrics. Several similarity metrics are proposed that outperform existing ones in the sense of attaining significantly higher Pearson and Spearman-rank correlations with the ground truth. The new metrics include Adjusted Percentage Error, Bilevel Gradient Histogram, Connected Components Comparison and combinations of such.

Another portion of the dissertation focuses on the aforementioned MTC, which is a block-based image coder that uses texture similarity metrics to decide if blocks of the image can be encoded by pointing to perceptually similar ones in the already coded region. The key to its success is an effective texture similarity metric, such as an LRI-based metric, and an effective search strategy. Compared to traditional image compression algorithms, e.g., JPEG, MTC achieves similar coding rate with higher reconstruction quality. And the advantage of MTC becomes larger as coding rate decreases.