摘要
立体图像的景物生动逼真,给人一种身临其境的全新视觉享受,但在制作、存储和传输过程中往往会产生失真。为了评价立体图像的质量优劣,提出了一种基于轮廓波(Contourlet)变换的无参考立体图像质量评价算法。通过对失真的左、右图像分别进行主成分分析(PCA)融合来生成新的融合图像,并使用基于SSIM(Structural Similarity)立体匹配算法生成视差图和匹配差值图,然后对上述三张图片进行Contourlet变换,再然后使用自定义的高频能量指标并结合边缘强度和信息熵,最后将得到的特征输入支持向量回归(Support Vector Regression,SVR)模型中学习,得出质量评价分数。该方法在德克萨斯大学公布的立体图像库中进行了验证,线性相关系数和斯皮尔曼相关系数在Phase I库中可高达0.957和0.947,在Phase II库中也可高达0.944和0.934,与主观评价吻合度很高,优于最新的一些评价方法。
Stereo images are vivid and lifelike, which give people a new visual enjoyment, but in the process of production, storage and transmission the distortion is often introduced. In order to evaluate the quality of stereo images, a new algorithm for quality evaluation of no-reference stereo images is presented, which is based on Contourlet transform. First, PCA image fusion is used to deal with the distortion image pairs to generate 2D images, and the Disparity Maps and Matching Difference Maps are generated by stereo matching algorithm based on SSIM（Structural Similarity）. Then, the amplitude of Fourier transform, the average gradient, information entropy as features are extracted. Finally, the extracted features are put into the SVR（Support Vector Regression） model to predict image quality scor~ The method has been verified on the LIVE 3D Image Quality Database which is published by university of Texas respectively. The linear correlation coefficient and spearman rank order correlation coefficient can be as high as 0. 957 and 0. 947 respectively in Phase I database, and 0. 944 and 0. 934 in Phase II database. The results have a high degree of agreement with the subjective evaluation and are better than some of the latest evaluation methods.
作者
李永生
桑庆兵
LI Yongsheng SANG Qingbing(School of Intemet of Things Engineering, Jiangnan University, Wuxi 214122, Jiangsu, China)
出处
《光学技术》
CAS
CSCD
北大核心
2016年第6期538-544,共7页
Optical Technique
基金
国家自然科学基金(61170120)
江苏省产学研前瞻性联合研究项目(BY2013015-41)
关键词
SSIM立体匹配
CONTOURLET变换
无参考立体图像质量评价
支持向量回归
主成分分析图像融合
stereo matching of SSIM
Contourlet transform
no-reference quality assessment for stereoscopic images
support vector regression（SVR）
principal component analysis（PCA） image fusion
