Just Noticeable Difference Prediction and Image Quality Assessment

Abstract

The paper shows that the information of the first just noticeable difference (JND) point position can significantly improve the performance of the objective peak signal-to-noise ratio (PSNR) measure in assessing the quality of JPEG compressed images. The degree of improvement depends on the choice of the first JND point position prediction model. Also, the paper shows that simple features derived from the gradient magnitude (spatial information and spatial frequency) of the original uncompressed image can be used for reliable position prediction. The analysis was conducted on two publicly available JND subject-rated image datasets MCL-JCI and JND-Pano. Among others, the linear correlation coefficient is used as an objective measurement parameter in prediction and in image quality assessment analysis. The prediction based on spatial frequency provided the best results, with over 95% of agreement with ground truth JND points position. This simple picture-wise prediction model has significantly improved the performance of conventional PSNR measure, with over 90% of agreement with subjective scores in image quality assessment. The PSNR performance is most enhanced by using a deep learning approach, where the correlation with subjective test results is close to 92%.