No-reference quality assessment for contrast-distorted images

Abstract

Contrast distortion is a common distortion type in the image applications. However, there are still very limited approaches proposed for quantifying the quality of the contrast-distorted images reliably. In this paper, we devise a novel no-reference/blind quality assessment method for those contrast-distorted images. In the proposed method, we characterize the image quality by deeply investigating multiple contrast distortion-relevant properties of the image, i.e., spatial characteristics, image histogram, visual perception characteristics and chrominance, which can describe the image quality more comprehensively and precisely. Accordingly, a series of quality-aware features are developed to characterize the contrast-distorted image quality properly. Support vector regression (SVR) is then employed to integrate all the extracted features and infer the image quality score. Extensive experiments conducted on the standard contrast-distorted image databases/datasets demonstrate that the proposed method achieves superior prediction performance to the state-of-the-art NR quality assessment models on evaluating the contrast-distorted image quality.