互信息域中的无参考图像质量评价

目的无参考图像质量评价是近几年来的研究热点,具有深远的现实意义和广泛的应用价值,提出一种基于互信息的无参考图像质量评价方法。方法该方法使用原始自然图像及其对应的规范化亮度图像和局部标准差图像作为输入,利用自相关互信息对输入图像邻近像素间的相关性进行量化,并引入多尺度分析得到图像在两个尺度上的互信息特征,最后使用支持向量机(SVM)在LIVE图像数据库上训练学习,从而对多类失真图像进行客观质量评价。结果在LIVE图像数据库中对本文算法进行性能测试,实验结果显示该算法得到的评价结果与人眼主观评价结果之间的平均相关系数高达0.93,总体分类准确率达到79%,性能足以与当前主流的全参考、无参考方法相竞争。结论本文方法有别于传统的基于变换的无参考图像质量评价方法,将着眼点放于自然图像邻近像素之间的固有联系上,并取得了较好的实验效果。由于没有使用图像变换并从全局域进行考虑,本文方法具有较低的时间复杂度。

No-reference image quality assessment in mutual information domain

Objective: As research hotspot in recent years, no-reference image quality assessment has profound practical significance and broad application value. We present a new method of no-reference (NR) image quality assessment (IQA) based on mutual information. Method: Original natural image and their corresponding normalized luminance field and local standard deviation field are used as inputs. Self-correlated mutual information is used to quantify the correlations between neighboring pixels of 3 categories of inputs, and the quantization results are used as features. In addition, the multiscale analysis is introduced to obtain the mutual information features across two scales. The image distortion classifier and quality prediction model are trained by using a support vector machine (SVM) on the LIVE image database and conduct the NR IQA across multiple categories of distortions. Result: We conduct the performance evaluation for our proposed algorithm on the LIVE image database, the experimental results show that the mean correlation coefficient between the quality judgment of this algorithm and the human subjective quality judgment is up to 0.93, and the total classification accuracy is up to 79%, delivering a performance which is competitive with the most popular full-reference (FR) / NR IQA methods. Conclusion: The method presented is different from the traditional NR IQA methods based on image transforms. Since natural image is highly structured, we focus on the inherent correlations between neighboring pixels of natural image, rather than the distribution of transformed coefficients, and obtain a good performance. Since the method presented is build without any image transforms and it is a global method, it has a relatively low time complexity.