An Explainable Spectral Analysis For Light Field Image Quality Assessment

Abstract

Translate Article

The light field (LF) technology is a promising way to achieve the immersive multimedia. The LF image quality assessment is vital for user experience. Compared with traditional twodimensional images, the LF images suffer from not only spatial distortion but also angular distortion, which destroys the light field angular consistency. Many algorithms are proposed to handle the objective light field image quality assessment. However, there is no mathematical description for the LF angular consistency and no discussion of spectral analysis in the angular domain. In this paper we provide an explainable spectral analysis for the LF image quality assessment. First, we formulate the LF as a 4 D signal. It is well-known that the energy of the light field Epipolar Plane Image (EPI) in the frequency domain should concentrate in a double-wedge area, but for the first time we introduce it into the LFI quality assessment. We test and discuss how the different distortions affect the energy distribution. The spectral features for blind LF image quality assessment are further proposed to complement the spatial features extracted by off-the-shelf algorithm. Experimentally we demonstrate that the features based on spectral energy distribution are more sensitive to angular distortion. Consequently, it can be a good guidance for a better LF quality assessment algorithm design.