\r李素梅¹，王明毅¹，赵 平¹，秦龙斌\r^{1, 2}\r
\r
AuthorsHTML:\r李素梅¹，王明毅¹，赵 平¹，秦龙斌\r^{1, 2}\r
\r
AuthorsListE:\rLi Sumei¹，Wang Mingyi¹，Zhao Ping¹，Qin Longbin\r^{1, 2}\r
\r
AuthorsHTMLE:\rLi Sumei¹，Wang Mingyi¹，Zhao Ping¹，Qin Longbin\r^{1, 2}\r
\r
Unit:\r\r1. 天津大学电气自动化与信息工程学院，天津 300072；\r
\r\r2. 昌都市公安局，昌都854000\r
\r
\r
Unit_EngLish:\r1. School of Electrical and Information Engineering，Tianjin University，Tianjin 300072，China；
2. Changdu Public Security Bureau，Changdu 854000，China\r
\r
Abstract_Chinese:\r\r\r\r本文基于深度卷积神经网络和融合图像提出了一种引入投影权值归一化的立体图像质量评价方法．首先基于人眼双目竞争现象，提出对经过\r\rGabor\r滤波后的左右视点图像进行彩色融合，从而得到单幅融合图像．卷积神经网络的输入即为预处理后的融合图像，通过卷积层自主对图像特征进行提取，采用池化层对特征信息降维，保留显著特征且减小网络计算复杂度；采用\rReLU\r非线性激活函数缓解梯度消失，有效缓解了网络过拟合问题；网络引入数据批量归一化来规范各层输入数据的分布，引入投影权值归一化来保证权值参数的量级相同，有效地提升了算法的性能．本文在公开的立体图像库\rLIVE\r-\rⅠ和\rLIVE\r-\rⅡ上进行了实验．实验结果表明，本文方法在对称失真与非对称失真的立体图像质量评价上均具有较好的性能．\r\r\r\r
\r
Abstract_English:\r\rCurrently\r，\rthe wide application of three-dimensional\r(\r3D\r)\rimagery makes stereoscopic image quality assessment\r(\rSIQA\r)\rincreasingly important in industry. Extensive application of deep learning promotes the development of SIQA methods based on convolution neural networks\r(\rCNN\r)\r. Compared with the traditional hand-crafted feature extraction process\r，\rCNN conforms more to human brain mechanisms\r．\rHowever\r，\rwith the depth of CNN\r，\roptimization of parameters and data in CNN is essential to improve the performance and efficiency of a network. In addition\r，\rby simulating the process of stereoscopic images in the human brain\r，\rthe construction of image fusion methods consistent with the human visual system has become an issue in SIQA. Based on deep CNN\r，\ra SIQA method is proposed using projection-based weight normalization. This method implemented a Gabor filter to left and right images\r，\rand fused them to obtain one new\r，\rcolorful\r，\rcyclopean image. Model input was the fused image after pretreatment. Convolution layers were applied to extract features autonomously\r，\rand pooling layers were adopted to reduce the dimension of feature information\r，\rretain significant features\r，\rand reduce the computational complexity of the network. Rectified linear units\r(\rReLU\r)\rwere used to mitigate the disappearance of gradients\r，\reffectively alleviating over-fitting. In addition\r，\reach layer’s inputs were identically distributed using batch normalization. Projection-based weight normalization ensured that the weight of each neuron had almost the same magnitude. These methods improved the model’s performance. Experiments were conducted on the published LIVE-\rⅠ\r and LIVE-\rⅡ\r databases. Experimental results showed that the method performs well on both symmetrical and asymmetrical stereoscopic image databases.\r\r
\r
Keyword_Chinese:立体图像质量评价；卷积神经网络；投影权值归一化；数据批量归一化\r

Keywords_English:stereoscopic image quality assessment；convolution neural network(CNN)；projection-based weight normalization(PBWN)；batch normalization(BN)\r


PDF全文下载地址:http://xbzrb.tju.edu.cn/#/digest?ArticleID=6419