Minimum image quality assessment based on saliency maps: a human visual approach

Image quality assessment as perceived by humans is of crucial importance in numerous fields of image processing. Transmission and storage of digital media require efficient methods to reduce the large number of bits to store an image, while maintaining sufficiently high quality compared to the original image. Since subjective evaluations cannot be performed in various scenarios, it is necessary to have objective metrics that predict image quality consistent with human perception. However, objective metrics that considers high levels of the human visual system are still limited. In this paper, we investigate the possibility of automatically predict, based on saliency maps, the minimum image quality threshold from which humans can perceive the elements on a compressed image. We conducted a series of experimental subjective tests where human observers have been exposed to compressed images with decreasing compression rates. To measure the difference between the saliency maps of the compressed and the original image it was used the normalized absolute error metric. Our results indicate that the elements on the image are only perceived by most of the human subjects not at a specific compressed image quality level, but depending on a saliency map difference threshold.