No-reference stereo image quality evaluation method integrating monocular and binocular features
Jinxin Bai, Xiaojuan Hu · 2023
In recent years, the widespread adoption of 3D technology has created a need for accurate and efficient stereoscopic image quality assessment methods. Such methods aim to automatically supervise and optimize stereoscopic image and video processing systems. Based on the characteristics of human vision, we propose a method that integrates monocular and binocular features. Initially, monocular quality-sensitive features are extracted separately for the left and right views through two parallel CNN sub-networks. Subsequently, a third CNN sub-network extracts binocular quality-sensitive features for stereo image pairs. Quality scores are then computed based on both monocular and binocular features, and the final predicted quality score is obtained through quality fusion. The proposed method is validated on the Waterloo dataset, achieving a PLCC value of 0.979 and SROCC value of 0.975 on Waterloo P-I, and a PLCC value of 0.973 and SROCC value of 0.970 on Waterloo P-II. The results demonstrate that the predicted scores obtained by the proposed method exhibit excellent consistency with subjective evaluation scores.