Variable selection for image quality assessment using a Neural Network based approach

Compressed image quality assessment is of increasing importance in image coding systems where the schemes optimization is based on the distortion measure. There exist many distortion measures in the literature which are often validated by comparing them to the human appreciation of the image quality, in particular the Mean Opinion Score (MOS). Until now, we do not know precisely which factors intervene into the human evaluation of the image quality. In this paper, we attempt to answer this question. We study a set of indicators and see what are the most relevant for the image quality assessment by using an Artificial Neural Network based model. The variable selection system results in defining the image indicators that convey relevant information for the subjective evaluation of image quality.