基于双重注意力和分层感知表征的IQA方法  

Image Quality Assessment Method Based on Dual Attention and Hierarchical Perceptual Representation

作　　者：史再峰 佟博文 孔凡宁 康泰 罗韬[3] Shi Zaifeng;Tong Bowen;Kong Fanning;Kang Tai;Luo Tao(School of Microelectronics,Tianjin University,Tianjin 300072,China;Tianjin Key Laboratory of Imaging and Sensing Microelectronic Technology,Tianjin 300072,China;College of Intelligence and Computing,Tianjin University,Tianjin 300072,China)

机构地区：[1]天津大学微电子学院,天津300072 [2]天津市成像与感知微电子技术重点实验室,天津300072 [3]天津大学智能与计算学部,天津300072

出　　处：《天津大学学报（自然科学与工程技术版）》2024年第3期234-243,共10页Journal of Tianjin University：Science and Technology

基　　金：国家自然科学基金资助项目(62071326);天津市科技计划资助项目(22JCYBJC00140).

摘　　要：图像质量评价在图像处理领域有着广泛的应用.基于深度学习的方法以多通道特征的形式获取图像信息,但在特征下采样过程中会丢失局部空间细节,导致对于图像信息的表征能力不足.针对以上问题,本文基于人类视觉系统的分层感知机制,提出了一种全参考图像质量评价方法,采用Siamese结构的卷积神经网络实现非线性映射,从不同尺度和视觉复杂度表征图像信息,并通过双重注意力模拟人类在评价图像质量时对视觉注意力的调控过程.此方法在特征提取阶段引入空间注意力机制,对特征图的二维空间位置赋以权重,计算人在感知图像失真信息时对空间区域的注意力差别.在特征融合阶段利用分组通道注意力模块显式建模通道间的依赖关系,对感知差异特征进行自适应的校准,使网络关注对于图像质量评价影响大的通道特征.实验结果表明,该方法在LIVE、TID2013和CSIQ 3个公开数据集上的斯皮尔曼相关系数分别达到0.975、0.938和0.963,在应对复杂失真类型图像时的性能提升显著,与人类主观评价的一致性良好.Image quality assessment(IQA)has extensive applications in the field of image processing.Deep learningbased methods are used to acquire image information in the form of multi-channel features;however,the downsampling process in the feature extraction stage can result in the loss of local spatial details,leading to an incomplete representation of image information.To address these issues,this study proposes a full-reference IQA method based on the hierarchical perception mechanism used in the human visual system.A Siamese-structure convolutional neural network is used to achieve nonlinear mapping,which represents image information based on different scales and visual complexities.The proposed method simulates the process of visual attention regulation used in humans for evaluating image quality through dual attention mechanisms.In the feature extraction stage,a spatial attention mechanism is introduced to assign weights to the two-dimensional spatial positions of the feature maps,modeling the attention differences in the human perception of distorted image information across different spatial regions.In the feature fusion stage,a group channel attention mechanism is used to explicitly model the dependency relations among channels,adaptively calibrating the perceived differential features and enabling the network to focus on channel features having a greater impact on IQA.Experimental results demonstrate that the proposed method achieves a Spearman correlation coefficient of 0.975,0.938,and 0.963 on the publicly available datasets LIVE,TID2013,and CSIQ,respectively.The proposed method substantially improves quality assessment performance when dealing with complex distortion types occurring in images and shows good consistency with human subjective evaluation.

关 键 词：图像质量评价 卷积神经网络 双重注意力机制 分层感知表征 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]