\r李素梅¹，薛建伟¹，秦龙斌\r^{1, 2}\r
\r
AuthorsHTML:\r李素梅¹，薛建伟¹，秦龙斌\r^{1, 2}\r
\r
AuthorsListE:\rLi Sumei¹，Xue Jianwei¹，Qin Longbin\r^{1, 2}\r
\r
AuthorsHTMLE:\rLi Sumei¹，Xue Jianwei¹，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长时间观看立体图像会导致视觉疲劳、恶心、头痛等不适感，如何结合人类的视觉特性对立体图像质量进行评价是近年来立体成像领域的研究热点．为此，本文提出了一种基于融合图像的无参考立体图像质量评价方
法．为了更好地模拟人脑处理立体图像的过程，提出了一种先将左、右视图融合然后进行处理的融合图像算法．首先针对左、右视图分别在RGB 3个通道上进行Gabor 滤波以模拟人眼的视觉多通道特性，获取其不同尺度和方向的结构特征，随后通过对比敏感度函数滤除图像的不重要频率信息，然后通过增益控制原理进行加权获得融合图像．相比之前无参考立体图像质量评价方法局限于提取手工裁剪的特征，本文采用直接将原始图像进行切块后，送入网络进行训练，让卷积神经网络自动地提取图像的特征，并且采用重叠切块的方法，相比于非重叠切块，重叠切块可以更好地保留相邻像素之间的关系，并且增加了训练的数据集．然而，与拥有几百万张图像组成的Imagenet 数据库相比，立体图像数据库仅有几百张，而且构成图像的基向量是普适的，所以本文对在Imagenet 数据库上训练好的Alexnet 网络进行迁移学习，建立输入图像和输出质量值之间端到端的映射．迁移学习网络模型较传统卷积神经网络收敛快，而且具有更好的初始权重．最后，鉴于人眼在观看图像时总是倾向于从图像的中心开始寻找视觉注视点，然后其注意力由中央向四周递减，本文利用显著特性对图像小块的输出进行加权以更好地模拟人眼的视觉显著特性．在公开的LIVE3D phase-Ⅰ、LIVE3D phase-Ⅱ数据库上进行测试，结果表明本文所提方法在对称和非对称立体图像数据库上较其他方法均取得了较好的结果，能够与人类的主观感知保持良好的一致性．\r
\r
Abstract_English:\rViewing stereoscopic images for a long time can lead to discomfort such as visual fatigue，nausea，and headaches．How to evaluate the stereoscopic image quality by combining human visual characteristics is a research hotspot in the field of stereoscopic imaging in recent years．For this reason，this paper proposes a non-reference stereoscopic image quality assessment method based on cyclopean images．In order to better simulate the process of the human brain processing stereoscopic images，a cyclopean image algorithm is proposed，in which the left and right views are firstly fused and then processed．The paper firstly performs Gabor filtering on the three channels of RGB for the left and right views to simulate the visual multi-channel characteristics of the human eye，and obtains the structural features of different scales and directions．Then，the unimportant frequency information of the image is filtered by contrast sensitivity function．And then，the cyclopean image is weighted by gain control principle．Previously，no-reference stereoscopic image quality assessment algorithms have been limited to the extraction of reliable hand-crafted features based on an understanding of the insufficiently revealed human visual system．In this
paper，the original image is sliced directly and sent to the network for training，so that the convolutional neural network could automatically extract the features of the image．And the cyclopean images are overlapped and cut to serve as the input of the transfer learning network．Compared with non-overlapping segmentation，overlapping segmentation can better preserve the relationship between adjacent pixels and increase the training data set．However，compared with the Imagenet database，composed of millions of images，there are only a few hundred stereoscopic image databases，and the base vectors of the images are universal．So the Alexnet network trained on the Imagenet database is transfered in this paper to establish the end-to-end mapping between the input image and the output quality value．The transfer learning network model converges faster than traditional convolutional neural networks and has better initial weights．Finally，since the human eye always tends to look for visual gaze points from the center of the
image when viewing the image，then its attention decreases from the center to the periphery，we use weighted features to weight the image blocks to better simulate the visual salient features of human eyes．The paper is tested on the published LIVE3D phase-Ⅰ and LIVE3D phase-Ⅱ databases．The results show that the proposed method achieves better results than other methods both in symmetric and asymmetrical stereoscopic image databases，and keeps good consistency with human subjective perception．\r
\r
Keyword_Chinese:立体图像质量；Alexnet 网络；迁移学习；融合图像；视觉显著性\r

Keywords_English:stereoscopic image quality；Alexnet network；transfer learning；fusion image；visual saliency\r


PDF全文下载地址:http://xbzrb.tju.edu.cn/#/digest?ArticleID=6325