基于机器学习的无参考图像质量评价指标体系改进

Improvement of index system of no reference image quality assessment based on machine learning

为解决BRISQUE算法中单纯提取灰度空间指标特征的问题,进一步提升算法预测的准确性,文章在BRISQUE算法的基础上改进了无参考图像质量评价指标体系,提出了一种基于灰度空间和色彩空间的改进算法。同时,为了减少单一数据库造成的模型过拟合现象,提升算法的鲁棒性,该算法分别在LIVE和CSIQ数据库上分析了算法计算结果与DMOS值的相关性。实验结果表明,改进的BRISQUE算法评价结果与人类主观评价具有高度的一致性,较BRISQUE算法在一致性方面有一定程度的提升。

In order to solve the problem in BRISQUE that extracting index features only in gray space and further improve the accuracy of algorithm prediction,this paper improves the index system of no reference image quality evaluation.At the same time,in order to reduce the model overfitting caused by single database and improve the robustness of the algorithm.In this paper,the correlation between the algorithm results and DMOS values is analyzed on LIVE,CSIQ databases respectively.The experimental results show that the evaluation results of the improved BRISQUE algorithm are highly consistent with human subjective evaluation.There is a certain degree of improvement in consistency compared with the BRISQUE algorithm.