基于残差注意力与特征融合的无参考图像质量评价方法  

No-reference Image Quality Evaluation Method Based on Residual Attention and Feature Fusion

作　　者：闫嘉阔 司占军[2] YAN Jia-kuo;SI Zhan-jun(College of Light Industry Science and Engineering,Tianjin University of Science and Technology,Tianjin 300457,China;College of Artificial Intelligence,Tianjin University of Science and Technology,Tianjin 300457,China)

机构地区：[1]天津科技大学轻工科学与工程学院,天津300457 [2]天津科技大学人工智能学院,天津300457

出　　处：《印刷与数字媒体技术研究》2025年第2期56-64,共9页Printing and Digital Media Technology Study

摘　　要：目前基于深度学习的无参考图像质量评价方法存在图像语义特征提取不充分或图像语义感知不足等问题。因此,本研究提出了一种改进方法,利用残差注意力和特征融合来进行无参考图像质量评价。首先,使用Resnet-50网络提取图像的低阶特征,然后通过残差双注意力模块增强提取的特征,以有效学习关键信息并更新权重,从而提高对图像特征的关注度并提取高阶特征。使用语义感知网络生成相应的权值参数,通过特征融合模块将不同阶段的高低阶特征进行融合,以丰富高级特征图中不同层级之间的语义和位置信息,并捕捉局部和全局失真。最后,结合全局图像信息和局部图像信息,根据图像内容获取的相应参数,对图像质量进行预测。实验结果表明,本研究提出的方法在预测失真图像质量时能够充分挖掘失真图像的高级语义信息,有效解决真实失真场景下失真图像内容变化和未知失真类型复杂的问题。在复杂的失真情况下,本研究方法仍能准确评价图像质量,并具有较强的泛化能力。Currently,there are some issues with deep learning-based no-reference image quality evaluation methods,such as insufficient extraction of semantic features or inadequate perception of image semantics.Therefore,an improved method for no-reference image quality evaluation using residual attention and feature fusion was proposed in this study.Firstly,the low-level features of the image were extracted using the Resnet-50 network.Then,the extracted features were enhanced using a residual dual-attention module to effectively learn key information and update weights,thereby improving the focus on high-level features and extracting high-order features.The semantic perception network was utilized to generate corresponding weight parameters,and the feature fusion module was used to fuse the high and low-level features from different stages,enriched the semantic and positional information between different levels in the high-level feature maps,and capturing local and global distortions.Finally,by combining global and local image information and using the obtained parameters based on image content,the image quality was predicted.The results showed that the proposed method can fully exploit the high-level semantic information of distorted images for accurately predicting the quality of distorted images,effectively addressing the challenges of content variations and unknown distortion types in real-world distortion scenarios.Even in complex distortion conditions,this method can accurately assess image quality and exhibits strong generalization capabilities.

关 键 词：无参考图像质量评价 残差注意力 特征融合 语义感知 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TS8[自动化与计算机技术—计算机科学与技术]