基于纹理奇异值分解的全参考图像质量评价  

Full-reference image quality assessment based on texture singular value decomposition

作　　者：李佳欣 段发阶[1] 傅骁[1] 牛广越 LI Jiaxin;DUAN Fajie;FU Xiao;NIU Guangyue(State Key Laboratory of Precision Measuring Technology&Instruments,Tianjin University,Tianjin 300072,China)

机构地区：[1]天津大学精密测试技术及仪器国家重点实验室,天津300072

出　　处：《光学精密工程》2025年第1期107-122,共16页Optics and Precision Engineering

基　　金：国家自然科学基金资助项目(No.52205573,No.U2241265,No.92360306,No.61971307,No.62231011);中国博士后科学基金资助项目(No.2022M720106);天津大学科技创新领军人才培育“启明计划”项目(No.2024XQM-0012);精密测试技术及仪器全国重点实验室(天津大学)青年教师科研启动项目(No.Pilq2304);国家科技重大专项(No.J2022-V-0005-0031);航空科学基金资助项目(No.2022Z060048001);青年人才托举工程资助项目(No.2021QNRC001);装备预研教育部联合基金资助(No.8091B022144);国防科技重点实验室基金资助项目(No.6142212210304);广东省重点研发计划项目(No.2020B0404030001);霍英东教育基金会资助项目(No.171055)。

摘　　要：对于工业领域的视觉系统,主观评价成本高,无参考图像质量评价预训练耗时长,需要高准确度的全参考图像质量评价模型。为此,提出了一种基于纹理信息加权的奇异值分解全参考图像质量评价模型。对参考图像块进行奇异值分解,利用参考图像块的奇异向量与失真图像块估计失真块的奇异值,据此得到亮度相似度分量;利用估计的失真图像块的奇异值评价平均偏移失真与对比度变化失真,得到对比度性相似度分量;通过失真图像块与参考图像块的奇异向量对单位矩阵的偏移量估计图像的结构相似度;最后,利用图像的纹理信息进行对亮度、对比度、结果相似度进行加权,得到全参考图像质量评价模型。依据4项评价标准,在6个常用的图像质量评价数据库上进行测试。实验结果表明,本模型在上述数据集中的加权Spearman秩序相关系数为0.8963;对于对比度变化失真,本模型Spearman秩序相关系数为0.8595,比第二名提高了85%。与多种全参考图像质量评价模型相比,本模型在预测精度、泛化性与计算复杂度方面具有明显的优势。In industrial vision systems,subjective assessment is costly,pre-training for no-reference quality evaluation is time-intensive,and there is a critical need for a highly accurate full-reference image quality assessment model.To address these challenges,this study proposes a novel full-reference image quality assessment model based on singular value decomposition(SVD)with weighted texture information.First,SVD is applied to the reference image blocks,and the singular values of the distorted blocks are esti-mated using the singular vectors of both the reference and distorted image blocks,yielding the brightness similarity component.Next,the estimated singular values of the distorted image blocks are used to quantify average offset distortion and contrast change distortion,resulting in the contrast similarity component.The structural similarity of the images is then determined by analyzing the deviation of the singular vectors of the distorted image blocks from the unit matrix of the reference image blocks.Finally,the brightness,contrast,and structural similarity components are weighted using texture information to construct the full-reference image quality assessment model.The proposed method was evaluated on six widely used image quality assessment databases across four performance criteria.Experimental results demonstrate that the model achieves a weighted Spearman rank correlation coefficient of 0.8963 across the datasets.For contrast change distortion,the model attains a Spearman rank correlation coefficient of 0.8595,outperforming the second-best method by 85%.Compared to existing full-reference image quality assessment models,the proposed approach offers significant advantages in prediction accuracy,generalization capability,and computational efficiency.

关 键 词：图像质量评价 全参考 奇异值分解 纹理信息 图像对比度 

分 类 号：TP394.1[自动化与计算机技术—计算机应用技术] TH691.9[自动化与计算机技术—计算机科学与技术]