Blind light field image quality assessment with tensor color domain and 3D shearlet transform

Abstract

Blind light field image quality assessment (LFIQA) models are able to automatically evaluate the perceptual quality of the distorted light field image (LFI) under conditions where the original LFI is not available. Most existing blind LFIQA models predict the quality of the distorted LFI by separately measuring spatial quality and angular consistency, while ignoring the correlation between the two. This paper proposes a blind LFI quality assessment method based on tensor color domain and three-dimensional (3D) shearlet transform. The traditional color space conversion only handles color 2D images, so a color space conversion based on tensor decomposition is designed to extract the color information of the distorted LFI, that is the 5D tensor, from the tensor color domain, and three 4D tensor slices are obtained. Each slice is reshaped into 3D pseudo-video, and which are converted into 3D shearlet domain, respectively. 3D shearlet coefficients are able to represent the local spatial-angular information of the distorted LFI, and the perceptual features based on 3D shearlet transform are extracted for LFI quality prediction. Finally, the perceptual feature vector is fed into the support vector regression to predict the quality of the LFI. Extensive experimental results on three databases have proved the superiority of the proposed method, and it is competitive with the start-of-the-art 2D image quality assessment, video quality assessment and LFI quality assessment methods.