摘要： 立体图像质量评价（SIQA）是评估立体成像系统性能的一种有效方法。考虑到深度信息是立体图像的重要特征，提出一种结合卷积神经网络（CNN）与立体图像深度显著性特征的无参考SIQA方法。分别利用改进显著特征检测模型和高斯差分滤波器提取立体图像的显著特征和深度特征，并通过小波变换融合两者得到深度显著性特征。在此基础上，将深度显著性特征、对比度特征和亮度系数归一化特征作为输入特征对CNN进行模型训练，从而预测图像的质量分数。该方法在LIVE 3D IQA Phase Ⅰ、Phase Ⅱ、NBU 3D IQA图像库上的皮尔森线性相关系数分别为0.948、0.962、0.943，斯皮尔曼秩相关系数分别为0.937、0.961、0.902，在Phase II、NBU 3D IQA跨数据库上的斯皮尔曼秩相关系数分别为0.832、0.673。实验结果表明，该方法预测的质量分数符合人类主观感知，且具有较好的适用性和鲁棒性。

关键词: 立体图像质量评价, 多元特征, 立体显著性, 无参考图像评价, 卷积神经网络

Abstract: Stereoscopic Image Quality Assessment(SIQA) is an effective method to evaluate the performance of stereoscopic imaging systems.Considering that depth information is an important feature of stereoscopic images, an method of non-reference SIQA is proposed by combining a Convolutional Neural Network(CNN) and depth saliency features of stereoscopic images.The method employs the improved saliency feature detection model and the difference of Gaussian filter to extract the saliency features and depth features. Then the two kinds of features are fused by using wavelet transform to obtain depth saliency features.On this basis, the depth saliency features, contrast features and brightness coefficient normalization features are input into the CNN for training, and then the quality score of the image is predicted.The method is tested with stereoscopic images for quality evaluation in multiple databases, including LIVE 3D IQA PhaseⅠ, Phase Ⅱ, and NBU 3D IQA.Experimental results show that the proposed method exhibits a Pearson linear correlation coefficient of 0.948 on LIVE 3D IQA Phase I, 0.962 on LIVE 3D IQA Phase II, and 0.943 on NBU 3D IQA.Its Spearman rank correlation coefficients on the above three database are 0.937, 0.961, and 0.902 respectively.The experimental results on the Phase II and NBU 3D IQA cross-database show that the method exhibits a Spearman rank correlation coefficient of 0.832 and 0.673.The method can generate quality scores that conform to human perception, and displays high applicability and robustness.

Key words: Stereoscopic Image Quality Assessment(SIQA), multiple features, stereoscopic saliency, non-reference quality evaluation, Convolutional Netural Network(CNN)

中图分类号: 

• TP391.41