计算机工程

• •    

基于多元特征的立体图像质量评价方法

  

  • 发布日期:2020-12-08

Stereoscopic image quality assessment based on multiple features

  • Published:2020-12-08

摘要: 立体图像质量评价(Stereoscopic Image Quality Assessment,SIQA)是评估立体成像系统性 能的一种有效方法。卷积神经网络(CNN)在计算机视觉领域应用广泛,基于 CNN 的无参考立体图像质 量评价(NR-SIQA)方法层出不穷。考虑到深度信息是立体图像的重要特征,提出一种结合 CNN 与立体 图像深度显著性特征的 NR-SIQA:提取立体图像的深度显著性特征、对比度特征和亮度系数归一化特征作 为卷积神经网络的输入,对神经网络进行训练从而预测图像的质量分数。应用该算法对LIVE 3D IQA Phase I、Phase II、NBU 3D IQA 图像库中的立体图像进行质量评价,Pearson 线性相关系数分别为 0.948、 0.962、0.943,Spearman 秩相关系数分别为 0.937、0.961、0.902。对算法进行跨数据库测试,在 Phase II、NBU 3D IQA 图像库上的 spearman 秩相关系数分别为 0.832,0.637。实验结果表明算法预测的质 量分数符合人类主观感知,且具有较好的适用性和鲁棒性。

Abstract: 】 Stereoscopic Image Quality Assessment (SIQA) is an effective method to evaluate the performance of stereoscopic imaging systems. Convolutional neural network (CNN) is widely used in the field of computer vision, and CNN-based non-reference stereo image quality assessment (NR-SIQA) methods are emerging in endlessly. Considering that depth information is an important feature of stereo images, a NR-SIQA combining CNN and stereo image depth saliency features is proposed: the depth saliency features, contrast features and brightness coefficient normalization features of stereo images are extracted as convolutional nerves The input of the network, the neural network is trained to predict the quality score of the image. The algorithm is used to evaluate the quality of stereo images in the LIVE 3D IQA Phase I, Phase II, and NBU 3D IQA image databases. The Pearson linear correlation coefficients are 0.948, 0.962, 0.943, and the Spearman rank correlation coefficients are 0.937, 0.961, and 0.902. The algorithm was tested across databases, and the spearman rank correlation coefficients on Phase II and NBU 3D IQA image databases were 0.832 and 0.637, respectively. Experimental results show that the qualityscore predicted by the algorithm conforms to human subjective perception, and has good applicability and robustness.