Dictionary and Structural Similarity-based Compression Artifact Reduction in DCT-coded Images

Abstract

We present an example and a structural similarity-based scheme to remove artifacts associated with block-based compression schemes. The proposed framework is designed to remove the compression artifacts on the basis of the training sets from the original and their compressed patches across block boundaries. For the data-learning phase, an image is first compressed using various strength levels. Next, each block boundary caused by the compression process is modeled and categorized into classes using a simple classifier. Then, we optimize the classes on the basis of the least mean square optimization approach. By contrast, for the artifact-removal phase, we classify a given block boundary into one of the classes, obtain the filter coefficients from the selected class, and apply these coefficients on two given blocks to remove the artifacts. In particular, the structural-similarity measurement is made only around the edges to effectively preserve the high-frequency image contents. The main advantages of the proposed algorithm are the following: (1) it is an efficient approach using only a fixed number of coefficients to remove blocking artifacts, (2) it requires no prior information on the blocking noise, and (3) it preserves the original image structure using structural similarity. To evaluate the proposed scheme, several state-of-the art approaches are described and compared in terms of peak signal-to-noise ratio, quality metric, and processing time.