A No-Reference Stereoscopic Image Quality Assessment Based on Cartoon Texture Decomposition and Human Visual System

It has become important to develop an objective stereoscopic image quality assessment (SIQA) method that aligns with human visual system characteristics. To enable an accurate and efficient assessment of stereoscopic image quality, this study introduces a no-reference stereoscopic image quality assessment model, aiming to address the limitations of existing assessment methods. Considering that natural images typically contain information like textures and contours, we decompose stereoscopic views into cartoon and texture images to effectively extract monocular perception features. We also take binocular difference information to explain binocular perception features. Subsequently, a CNN multi-branch architecture is employed to feed images into the network for extracting relevant feature mappings. Finally, all sub-networks are used for quality scoring predictions, resulting in the final perceptual quality score. Experiments conducted on the LIVE dataset have demonstrated the superiority of this approach.