Spatial-angular features based no-reference light field quality assessment

Light Field (LF) is capable of capturing light intensity as well as information about the direction and position of the light, resulting in an immersive visual experience. Visual quality can be greatly influenced by the distortion due to light field image (LFI) processing, such as compression or reconstruction. Therefore, the development of a perceptual quality measure for LFIs is imperative. Refocused images (RIs), as an important form of visualization applications of LF, capture the spatial quality, semantic and depth information, while epipolar plane images (EPIs) reflect the angular consistency. Based on these two kinds of views, a novel spatial- angular features based no-reference light field image quality assessment (NR-LFIQA) metric is presented in this work. Existing methods that use RIs to extract spatial features from LFIs typically follow approaches from the IQA field, neglecting the unique properties of RIs as multi-distorted images. Therefore, this paper proposes to analyze RIs by considering both edge and texture features. For angular features, this paper analyzes the shortcomings of LBP, and the variance-enhanced LBP (VELBP) is designed to improve the adaptability of LBP to light fields. Experimental results show that the proposed model outperforms the existing LFIQA models on four publicly available datasets. The source code for the paper will be published after it is organized.