Matched-Texture Coding and Metric Learning for Structurally Lossless Compression

Existing techniques rely on point-by-point metrics that cannot account for the stochastic and repetitive nature of textures. We propose a new compression approach, called matched-texture coding (MTC), that can exploit texture redundancy for dramatic increases in coding efficiency without any significant loss of visual quality. The new approach relies on new structural texture similarity metrics (STSIMs), and can be combined with traditional compression technicques (e.g., JPEG) so that it can be applied to any image. The main idea is to encode selected image blocks of textures by pointing to previously encoded blocks of similar textures. The same idea can be applied to smooth blocks and blocks that contain transitions between regions, smooth or textured. All blocks for which a good match cannot be found in the already encoded image region, are encoded with the baseline method. The proposed approach allows significant point-by-point differences between the original and the coded image. Such differences may be noticeable in a side-by-side comparison but do not affect the overall quality of the image, so that both could be considered original images. Indeed, the proposed approach represents the first coding algorithm that achieves structurally lossless compression. The key feature that distinguishes MTC from existing techniques, which rely on prediction to increase coding efficiency, like motion compensation and intra-block copy used in state-of-the-art video coders, in that there is no residual encoding. An image block can be simply replaced with a previously encoded block. To make this possible without loss of visual quality, MTC relies on structural texture similarity metrics, which account for the stochastic nature of textures and thus make it possible to exploit their redundancy (repetitiveness) and inherent masking ability. In addition to texture similarity metrics, the proposed approach incorporates texture blending techniques, to avoid stitching artifacts, and lighting correction techniques (low-frequency) adjustment of a texture patch), to avoid artifacts due to sudden changes in the local image average. Another key ingredient of the proposed approach is a side-matching criterion by which both the encoder and the decoder can select blocks from the already coded and reconstructed region that are potentially good candidates for encoding the next block with minimal side information. One of the key contributions of this thesis is the design both a general theoretical framework and a software implementation structure that can integrate all the key components of the proposed approach. In addition, we analyze the computation complexity of each component, in order to maximize implementation efficiency. The bottleneck for the success of MTC is the structural similarity metric. We consider the limitations of STSIM-2, one of the existing STSIMs, and propose new metrics to address them. In particular, we adopt the Mahalanobis version of STSIM, and use machine learning techniques to optimize its parameters. For that, we built a database of texture pairs, annotated with subjective similarity values. To construct the database we designed a subjective test that simulates MTC to obtain typical texture pairs and presents them to human observers. Experimental results with natural images demonstrate the advantages of the proposed approach.