Saliency-guided deep framework for image quality assessment

Abstract

Image quality assessment (IQA) has thrived for decades, and researchers continue to explore how the human brain perceives visual stimuli. Psychological evidence shows that humans prefer qualitative descriptions when evaluating image quality, yet most researches still concentrate on numerical descriptions. Furthermore, handcrafting features are widely used in this community, which constrains the models' flexibility. A novel model is proposed with two major advantages: the saliency-guided feature learning can learn features unsupervisedly, and the deep framework recasts IQA as a classification problem, analogous to human qualitative evaluation. Experiments validate the proposed model's effectiveness.