Assessing the Quality of Compressed Pictures Using a Perceptual Model

This paper presents an objective image quality metric based on a model of the human visual system. It essentially consists of two stages. Firstly, a multi-channel early vision model, tuned specifically for complex natural scenes, is used to determine the visibility of coding errors at each location in the scene. A model of visual attention is then used to identify regions of interest in the scene. This allows the visible errors to be weighted, depending on the perceptual importance of the region in which they occur. It is demonstrated that this technique produces a high correlation with subjective test data (0.93), compared to 0.87 when only the early vision model is used, and 0.65 for PSNR. A description of how this technique can be extended to video quality assessment is also included.

Published

1998-10

Content type

Original Research

DOI

10.5594/M00303

ISBN

978-1-61482-929-4