李素梅，丁义修，常永莉，韩 旭
AuthorsHTML:李素梅，丁义修，常永莉，韩 旭
AuthorsListE:Li Sumei，Ding Yixiu，Chang Yongli，Han Xu
AuthorsHTMLE:Li Sumei，Ding Yixiu，Chang Yongli，Han Xu
Unit:天津大学电气自动化与信息工程学院，天津 300072
Unit_EngLish:School of Electrical and Information Engineering，Tianjin University，Tianjin 300072，China
Abstract_Chinese:考虑到视差图在立体成像中的重要性，设计了一种双通道的卷积神经网络来实现无参考立体图像质量评价.首先，建立一个以密集连接网络为主体的卷积神经网络结构，用于提取特征. 其次，基于人类视觉系统的双目融合和双目竞争的特性，将左右视图进行R、G、B三通道融合得到彩色融合图像，并将此融合图像作为卷积神经网络的一个通道的输入；另一通道的输入为视差图，视差图起到了特征补偿的作用. 然后，通过改进挤压和激励模块来实现视差图对融合图像的加权指导. 这种加权策略加强了融合图像的重要信息的比重，减轻了非重要信息的比重. 最后，在卷积神经网络的末端，将视差图的特征和加权校正过的融合图像的特征进行融合得到总体特征，将总体特征与主观评价方法得分进行回归分析，得到待测立体图像的质量分数. 在两个公开的LIVE立体数据库上进行实验验证. 结果表明：所提出的无参考立体图像质量评价方法能够有效地应对对称和非对称失真类型的立体图像，并与主观评测方法保持高度一致.
Abstract_English:Given the importance of the disparity map in stereoscopic imaging，a two-channel convolution neural network(CNN)was designed to evaluate the quality of stereoscopic imaging without reference．Firstly，a CNN structure with a dense connection network as the main body was established for feature extraction．Secondly，a column of CNN input came from the cyclopean image，which was based on the characteristics of binocular combination and rivalry of the human visual system．Left and right views were fused into three channels to get the color cyclopean image．This image was used as the input of one channel of the CNN．The input of other channel was the disparity map，which provided some compensatory information for the cyclopean image．More importantly，we employed the features of the disparity map to guide and weigh the feature maps obtained from the cyclopean image，which were implemented by modifying the structure of the squeeze and excitation(SE)block．This weighting strategy strengthened the transmission of important information，and reduced the transmission of non-significant information from the cyclopean image．Finally，we combined the outputs from the two columns and processed them to get the final quality score of the stereoscopic image at the end of the CNN．The experiment was carried out on two open LIVE stereoscopic databases．Experimental results demonstrated that the proposed method could achieve highly consistent alignment with the subjective assessment．
Keyword_Chinese:立体图像质量评价；融合图像；卷积神经网络；视差图；挤压和激励模块
Keywords_English:stereoscopic image quality assessment；cyclopean image；convolution neural network；disparity map；squeeze and excitation block

PDF全文下载地址:http://xbzrb.tju.edu.cn/#/digest?ArticleID=6502