基于AHP的多特征融合图像质量评价算法  

Multi Feature Fusion Image Quality Evaluation Algorithm Based on AHP

作　　者：沈凡凡 刘海鹏 徐超[1] 陈勇 SHEN Fan-fan;LIU Hai-peng;XU Chao;CHEN Yong(School of Computer Science,Nanjing Audit University,Nanjing Jiangsu 211815,China)

机构地区：[1]南京审计大学计算机学院,江苏南京211815

出　　处：《计算机仿真》2024年第11期225-232,共8页Computer Simulation

基　　金：国家自然科学基金资助(61902189,71972102);江苏省高等学校基础科学(自然科学)研究项目资助(22KJA520004);江苏省教育科学“十四五”规划重点课题(C-b/2021/01/26);南京审计大学高教研究课题(2022JG002);江苏省研究生科研与实践创新计划项目(SJCX23_1100,SJCX22_1000);全国高等学校计算机教育研究会教育研究项目(CERACU2022R19)。

摘　　要：针对图像质量评价算法通用性不高的问题,在提取参考图像与失真图像的多种底层特征的基础上,计算各个特征在原图和失真图之间的相似度,结合AHP(Analytic Hierarchy Process)模型提出多特征融合的图像质量评价算法(Multi Feature Image Quality Assessment,MFIQA)。针对参考图像与失真图像分别提取图像特征并计算颜色、主体形状、局部特征和纹理细节相似度,根据图像类型和应用场景的不同,通过AHP建立对应的权重分配模型,将相似度数值归一化后代入模型中,最终得到量化的质量分数。在TID2008数据集上,该算法在KROCC上的表现相较于PSNR获得了7.3%的提升,相较于SSIM获得了3.5%的提升;在TID2013数据集上,该算法在RMSE上的表现相较于SSIM获得了17.9%的提升,相较于PSNR获得了5.4%的提升。在TID2008和TID2013数据集上的实验表明,文中算法的主客观一致性表现较好。In response to the problem of low universality of image quality evaluation algorithms,a Multi Feature Image Quality Assessment(MFIQA)algorithm is proposed by extracting multiple underlying features of reference and distorted images,calculating the similarity between each feature in the original and distorted images,and combining the Analytic Hierarchy Process(AHP)model.For the reference image and the distorted image,the image features are extracted and the similarity of color,body shape,local features and texture details is calculated.According to the different image types and application scenarios,the corresponding weight distribution model is established through AHP,and the similarity value is normalized into the model,and finally the quantized quality score is obtained.On the TID2008 dataset,the performance of the algorithm on KROCC has been improved by 7.3%compared with PSNR and 3.5%compared with SSIM;On the TID2013 dataset,the performance of the algorithm on RMSE has been improved by 17.9%compared with SSIM and 5.4%compared with PSNR.The experiments on TID2008 and TID2013 data sets show that the subjective and objective consistency of the algorithm in this paper is good.

关 键 词：图像质量评价 特征提取 多特征融合 全参考 视觉感知特性 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]