基于改进加权多源TrAdaBoost算法的无参考图像质量评价方法

Non-reference Image Quality Assessment Based on Improved Weighted Multisource TrAdaBoost

基于机器学习的无参考图像质量评价方法依赖于大量训练样本,但训练数据集的构建需要耗费大量人力物力。依据迁移学习理论,面向上述问题,首先提出了一种基于改进加权多源TrAdaBoost(weighted multisource TrAdaBoost,WMTrA)算法。算法的无参考图像质量评价算法采用权重自动更新方式,挖掘辅助图像库中的有价值样本,只需少量目标图像库样本便可以建立准确的图像质量评价模型;然后,将它应用到无参考图像质量评价方法上,检测了其效果。在JPEG,JPEG2000失真图像上的评价结果表明,相比于传统机器学习算法,本文方法能够有效利用现有辅助数据集,减少对目标数据集的数量要求,是一种具有实用性的无参考图像质量评价方法。

Non-reference(NR)image quality assessment(IQA)based on machine learning(ML)need a large number of training samples,which costs a lot of manpower and resources.Based on classical transfer learning(TL)theory,aiming at NR-IQA work s real problem,research divided into two parts,firstly,an improved weighted multisource TrAdaBoost(WMTrA)algorithm is proposed,which utilize automatic weight update model,mine valuable samples in the auxiliary samples,so it can establish an accurate image quality assessment model by a few samples of the target image samples.Secondly,the improved WMTrAdaBoost algorithm is adopted to do NR-IQA work,to test its real performance.Results on the JPEG and JPEG2000 distorted images show that,compared with the traditional ML algorithm,the proposed algorithm can unitize existing auxiliary data sets effectively and reduce number of the target data sets,and it is a practical NR-IQA algorithm.