基于自然统计特征分布的无参考图像质量评价

A No-reference Image Quality Assessment Based on Distribution Characteristics of Natural Statistics

针对目前的无参考评价方法无法准确反映人类对图像质量的视觉感知效果,该文提出一种基于自然统计特征分布(DIstribution Characteristics of Natural statistics,DICN)的无参考图像质量评价方法。其原理是用小波变换将图像分解为低频子带和高频子带部分,再将高频子带部分分成8′8的小块,提取每一子块的幅值和信息熵,并分别计算其分布直方图均值和斜度作为特征,利用支持向量回归思想对特征进行训练,建立5种不同失真类型的质量预测模型。在此基础上,采用支持向量机针对图像特征构造分类器并进行失真判断以确定不同失真的权重,结合5种失真评价模型可得到自然统计特征分布的无参考评价模型。实验结果分析表明,该算法的评价效果优于现有的经典算法,与主观评价具有较好一致性,能够准确反映人类对图像质量的视觉感知效果。

The current No-Reference Image Quality Assessment（NR-IQA） methods are not well consistent with subjective evaluation, a novel NR-IQA method based on the DIstribution Characteristics of Natural statistics（DICN） is proposed in this paper. In the proposed method, image is decomposed into low frequency subbands and high frequency subbands with wavelet, and its high frequency subbands are divided into blocks at size of 8×8, their amplitude and entropy are respectively extracted from the blocks, then their mean values of the distribution histogram and skewness are respectively calculated, and their results are as the image features. The features trained by Support Vector Regression（SVR） are for building 5 kinds of distortion image quality pre-measurement model. To determine the weights of the different distortions, the image features of classifier based on SVR are structured for carrying out the distortion evalution. Based on 5 kinds of distortion evaluation models, the NR-IQA model with the natural statistical distribution can be obtained. The results of experiments show that the proposed method performance is better than the present classical methods. The method is well consistent with the subjective assessment results, and can reflect human subjective feeling well.