基于特征优化的无参考光场图像质量评价

No-Reference Light Field Image Quality Assessment Based on Feature Optimization

光场图像因其能够捕捉光的方向信息而备受学术界和产业界的广泛关注,然而,光场图像在压缩和重建过程中常会出现不同程度的失真,影响光场图像的后续处理和应用。因此,需要设计一个光场图像质量评估器来估计失真光场图像的质量。传统的光场图像评价方法在提取光场图像的空间特征和角度特征时,未充分考虑人眼视觉的多通道特性以及人眼对角度变化的敏感性,从而影响光场图像的质量评价结果。为此,提出一种无参考光场图像质量评价方法。设计多频带局部二值模式算法,提取光场图像的空间特征并利用优化提取的空间特征测量光场图像的空间质量。提出加权局部相位量化算法,该算法在对微透镜图像单元中提取的角度特征进行相位量化时,根据角度信息变化的强弱赋予不同的权值。在此基础上,将空间和角度纹理特征结合成一维特征向量,输入到已经训练的支持向量回归中,得到光场图像的质量分数。在Win5-LID和NBU-LF1.0数据集上的实验结果表明,该方法的斯皮尔曼等级相关系数分别为0.799 1和0.803 2,相比SSIM、FSIM、BRISQUE等方法,具有更优的质量评估准确性和稳定性。

Light Field Image(LFI)is receiving widespread attention from academia and industry owing to its ability to capture directional information from light.However,LFI oftens experiences varying degrees of distortion during compression and reconstruction,which affects the subsequent processing and applications.Accordingly,an LFI quality evaluator can be used to estimate the quality of distorted LFI.Traditional LFI assessment methods do not fully consider the multi-channel characteristics of human vision,or the sensitivity of the human eye to angular changes,when extracting spatial and angular features,thereby affecting the quality assessment results.This paper proposes a method for assessing the quality of no-reference LFI.The multi-band Local Binary Pattern(LBP)algorithm is designed to optimize and extract the spatial features of LFI to measure the images'spatial quality.A weighted Local Phase Quantization(LPQ)algorithm is also proposed to assign different weights according to the strength of angular changes when performing phase quantization on angular features extracted from Micro-Lens Image(MLI)units.On this basis,the spatial and angular texture features are combined into one-dimensional feature vectors,which are input into the trained Support Vector Regression(SVR)to obtain the LFI's quality score.Experimental results on the Win5-LID and NBU-LF1.0 datasets show that the Spearman Rank-order Correlation Coefficient(SRCC)of the proposed method is 0.7991 and 0.8032,respectively.Compared to methods such as SSIM,FSIM,and BRISQUE,the proposed approach exhibits better quality assessment accuracy and stability.