深度视觉频谱残余融合的图像质量评价  

Image quality assessment based on pooling of deeply visual spectral residual

作　　者：丰明坤[1] 孙丽慧[1] 葛丁飞[1] 翟治年[1] 王海江 彭艳斌[1] FENG Mngkun;SUN Lihui;GE Dingfei;ZHAI Zhinian;WANG Haijiang;PENG Yanbin(School of Information and Electronic Engineering,Zhejiang University of Science and Technology,Hangzhou,Zhejiang 310012,China)

机构地区：[1]浙江科技学院信息与电子工程学院,浙江杭州310023

出　　处：《光电子．激光》2021年第10期1055-1064,共10页Journal of Optoelectronics·Laser

基　　金：浙江省公益性技术应用研究计划(LGF19F020005)资助项目。

摘　　要：针对现有图像质量评价方法的缺陷,通过深度学习理论建模人眼视觉系统(human vision system,HVS)特性,提出了一种基于视觉特征深度感知与学习融合(deeply perception and learning for pooling,DPLP)的评价方法。首先为了增加图像视觉特征的稳定性,根据人眼感光的空域结构特征和频域多通道特性,对图像依次进行二维Log-Gabor小波变换、梯度变换和频谱残余的深度视觉信息处理,然后分别提取各层视觉信息进行质量评价。其次为了克服HVS融合的不确定性,对质量评价信息采取了深度池化策略,第一层为评价视图的空域融合,采取了符合人眼感光特性的高斯加权策略;第二层为多通道评价的频域融合,采取了具有HVS推理能力的BP神经网络的学习-预测策略;第三层为各级视觉特征的评价融合,采取了具有自适应特性的回归函数策略。最后,基于现实中的各种失真类型图像进行了实验,结果表明所提方法具有较高的主客观评价一致性水平和更好的稳定性。To address the shortcomings of the existing image quality assessment(IQA)methods,an IQA method based on deeply perception and learning for pooling(DPLP)of visual features is proposed by deep learning theory modeling the characteristics of human vision system(HVS).Firstly,according to spatial structure feature and multi-channel characteristics of the human eye,deeply visual information of image is successively processed by two-dimensional Log-Gabor wavelet transform,gradient transform and spectrum residual,then each layer of vision information is extracted for quality assessment in order to increase the stability of image visual features.Secondly,the deeply pooling strategy is adopted for the quality assessment information in order to overcome the uncertainty of HVS pooling.The first layer is the spatial pooling of the assessment map and the Gauss weighting strategy which accords with the characteristics of human eye sensitivity is adopted.The second layer is the frequency pooling of multi-channel assessment and the learning-prediction strategy of BP neural network with HVS reasoning ability is adopted.The third layer is the assessment pooling of all levels of visual features and the regression function which holds adaptive characteristics is adopted.Finally,experiments are carried out based on various types of distortion images in reality.The results show that the proposed method has higher consistency level between subjective assessment and objective assessment and better stability.

关 键 词：图像质量评价 深度学习 池化策略 视觉频谱残余 

分 类 号：TN391.41[电子电信—物理电子学]