The strategy of image quality assessment: A new fidelity metric based upon distortion contrast decoupling.

摘要

This thesis presents a new image fidelity metric, Most Apparent Distortion (MAD), that uses a visual masking model and a measure of appearance distortion strategically to define the fidelity of a distorted image. Subjective image fidelity has been shown to be largely influenced by visual contrast masking of distortions and distortion energy. However, recent image fidelity metrics without an explicit visual masking model have been shown to correlate highly with subjective ratings. We argue that, at high quality, viewers use a different strategy for the task of rating distorted images than at low quality. We then evaluate the performance of MAD on two fidelity databases. In particular, we compare the performance of MAD to Peak Signal to Noise Ratio, Visual Signal to Noise Ratio, Structural Similarity, and Visual Information Fidelity. The results show that MAD performs statistically better than all other fidelity algorithms using various evaluation criteria for both databases.