Automatic prediction of saliency on JPEG distorted images

Abstract

We propose an algorithm to detect salient regions for JPEG distorted images for two tasks: quality assessment and free viewing. The algorithm extracts low-level features such as contrast, luminance, quality and so on and uses a machine-learning framework to predict salient regions in JPEG distorted images. We demonstrate that the automatically predicted regions-of-interest highly correlate with those from (human) ground truth saliency maps. Further, we evaluate the relevance of extracted low-level features for saliency prediction and analyze how incorporation of quality as a feature improves prediction performance as a function of the distortion severity. Applications of such a saliency prediction framework include developing novel pooling strategies for image quality assessment.