No-Reference Quality Assessment for Realistic Distorted Images by Color Moment and Texture Features
Ziliang Zhang, Yuming Fang, Jiebin Yan, Rengang Du · 2020
In real applications, images captured by cameras are easily affected by various distortion types, such as noise, blur, blockiness, and so forth. The complicated distortion existing in realistic distorted images is the major challenge to accurately predict their perceptual quality. In this work, we propose a novel and effective general-purpose No-Reference (NR) quality assessment method for realistic distorted images by Color Moment and Log-Gabor Layer. In the proposed method, we firstly convert the input image into HSV color space. Then, we extract the chromatic features namely color moment, which can capture the color degradations effectively. Meanwhile, considering the different sensitivities of the HVS to texture and non-texture regions, we divide the image into four layers by log-Gabor filter and extract the texture features in each layer. Quantitative experiments on two realistic distorted image datasets prove the validity of our method and its advantageous predicting image quality over the state-of-the-art models.