A Full-Reference Image Quality Assessment Model Based on Quadratic Gradient Magnitude and LOG Signal

摘要

Image quality assessment aims at estimating the subject quality of images and builds models to high efficiently evaluate the perceptual quality of the image for many applications. Because the human visual system (HVS) is highly sensitive to structural information, various image features have been studied and widely applied in IQA metrics design. Previous work has validated that the image gradient magnitude and the Laplacian of Gaussian (LOG) operator are efficient structural features in IQA tasks. Most of the IQA metrics work capably only when the distorted image is totally registered with the reference image, and perform poorly on images even with small translations. In this paper, we suggested an FR-IQA method with a simple combination of the gradient magnitude and the LOG signals, which obtains satisfied performance in evaluating image quality while considering the shift-invariance property for not well-registered reference and distortion image pair. Experimental results show that the proposed model works robustly on three large scale subjective IQA databases which contain a variety of distortion types and levels, stays in the state-of-the-art FR-IQA models and achieves the best performance in terms of weighted average score over the three databases. Furthermore, we proved that the proposed model performs better in translation-invariance test compared with the competitors.