基于视差信息的无参考立体图像质量评价

No-reference Stereo Image Quality Assessment Based on Disparity Information

近年来,随着深度学习在图像质量评价领域的快速发展,平面图像质量评价得到了有效的改善,但是立体图像质量评价还有待提高。为此,文中结合三分支卷积神经网络,提出了基于视差信息的无参考立体图像质量评价方法,并分析了不同视差图对模型性能的影响。该方法将左右视图以及视差图小块作为输入,自动提取特征,通过训练得到回归模型,从而实现对立体图像的预测。文中使用了5种不同立体匹配算法来生成视差图,实验结果表明使用SAD算法得到的效果最好。在立体图像库LIVE3D和MCL3D上的实验结果表明,该方法不仅适用于评估对称失真图像,还适用于非对称失真的立体图像评价。该方法在总体失真上的结果优于其他对比算法,尤其是在MCL3D图像库上,所提方法的PLCC和SROCC比其他方法高出1%和4%。实验数据表明,所提模型提高了立体图像质量评价的性能,与人类主观感知高度一致。

In recent years,with the rapid development of deep learning in the field of image quality assessment(IQA),2D-IQA has been improved,but 3D-IQA still needs to be improved.Therefore,combining the three-branch convolutional neural network,the paper proposes a no-reference stereo image quality assessment based on disparity information and analyzes the influence of different disparity maps on the performance of the model.The algorithm takes the left/right view patches and the disparity map patches as input,automatically extracts features,and obtains the regression model through training to realize the prediction of the stereo images.In this paper,5 different stereo matching algorithms are used to generate disparity maps.The experimental results show that the SAD algorithm is the best.The experimental results on stereo image databases LIVE3D and MCL3D show that the method is not only suitable for evaluating symmetric distortion images,but also for evaluating asymmetric distortion stereo images.The overall distortion results of this method are superior to other comparison algorithms.Especially on the MCL3D image database,the evaluation method PLCC and SROCC of the proposed method are 1%and 4%higher than other methods.The Experimental data shows that the proposed model improves the performance of stereo image quality assessment,which is highly consistent with human subjective perception.