李素梅，韩永甜，马 帅，韩 旭
AuthorsHTML:李素梅，韩永甜，马 帅，韩 旭
AuthorsListE:Li Sumei，Han Yongtian，Ma Shuai，Han Xu
AuthorsHTMLE:Li Sumei，Han Yongtian，Ma Shuai，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:立体图像质量评价为3D技术的发展与应用提供了技术支撑．如何根据立体图像特点，构建更加符合立体视觉认知机制的立体图像质量评价模型，已成为该领域的关键问题之一．现有的立体图像质量评价方法要么先分别处理左右视图，然后结合左右视图质量得到立体图像质量分数；要么先对左右视图进行融合得到融合视图，然后评价平面的融合视图得到立体图像质量分数．事实上，大脑对立体视觉信号的处理是一个长期的复杂融合与处理的过程，并最终在视觉皮层完成对视觉信号的认知与判断．受大脑立体视觉认知机制的启发，本文提出一种基于双目融合网络的立体图像质量评价模型，实现了双目信息的多次融合与处理，模拟了大脑对双目信息进行处理和判决的视觉传导通路．所提出的双目融合网络包含左视图通道、右视图通道和融合通道，模拟立体视觉信息在视觉通路中的逐层并行处理过程；左右视图通道在网络中多次交互，模拟视觉通路中双目信息的多次融合与处理；网络末端的3个全连接层，模拟视觉信息经过视觉通路处理后复杂的质量判断过程．本文实验在两个公开立体图像库LIVE 3D PhaseⅠ和LIVE 3D PhaseⅡ上进行，实验结果表明，该方法在对称与非对称失真立体图像上均能取得更好的结果，且较其他方法具有更强的普适性．
Abstract_English:Stereo image quality assessment provides technical support for the development and application of 3D technologies. The construction of a stereo image quality assessment model that is more in line with the stereo visual cognition mechanism based on the characteristics of stereo images has become one of the key issues in this field. Existing stereo image quality evaluation methods either process the left and right images respectively and then combine their quality to derive the stereo image quality score，or first fuse the left and right images to get the fusion image and then evaluate the planar fusion image to derive the stereo image quality score. In fact，the processing of stereo visual signals in the brain is a long-term process of complex fusion and processing，and finally completes the cognition and judgment of visual signals in the visual cortex. Inspired by the stereo visual transmission pathway in the brain，a stereo image quality assessment model based on the binocular fusion network is proposed in this study，which realizes multiple fusion and processing of binocular information，and simulates the brain’s visual pathway for processing and judging binocular information. The convolutional neural network model includes the left view channel，the right view channel and the fusion channel，which simulates the hierarchical and parallel processing of visual information in the visual pathway. The left and right view channels have multiple interactions in the network，which simulates the multiple fusion and processing in the visual pathway. Three fully connected layers are set at the end of the network，which simulates the complex quality judgment process after information fusion and processing in the visual pathway. Experiments are conducted on two public stereo image databases，namely LIVE 3D Phase I and LIVE 3D Phase II，and the experimental results show that the proposed method outperforms other methods in both symmetric and asymmetric distorted stereo images，and shows better universality.
Keyword_Chinese:图像质量评价；立体图像；立体视觉；双目融合
Keywords_English:image quality assessment；stereo image；stereo visual；binocular fusion

PDF全文下载地址:http://xbzrb.tju.edu.cn/#/digest?ArticleID=6533