基于深度学习的无参考立体图像质量评价

Blind Image Quality Assessment for Stereoscopic Images via Deep Learning

立体图像质量评价是评价立体视频系统性能的有效途径,而如何结合人类的视觉特性对立体图像质量进行评价是目前的研究难点.为此提出一种基于深度学习的无参考立体图像质量评价方法,分为训练和测试2个阶段.在训练阶段,首先对左右图像分别进行Gabor滤波,获取不同尺度和方向的统计特征作为单目特性;然后根据人眼视觉系统的双目竞争特性,将左右图像融合得到独眼图,提取其方向梯度直方图作为双目特征;最后通过深度信念网络训练得到特征和主观评价值之间的回归模型.在测试阶段,根据已建立的回归模型,预测得到左右图像质量并联合得到立体图像质量.实验结果表明,文中方法在对称和非对称立体图像数据库都取得了较好的效果,与人类的主观感知保持良好的一致性.

Stereoscopic image quality assessment is an effective way to evaluate the performance of stereoscopic video systems, but how to utilize human visual characteristics effectively is still a research focus in stereoscopic image quality assessment. In this paper, a blind image quality assessment method for stereoscopic images is pro-posed via deep learning. The proposed method is composed of two stages: training and testing. In the training stage, Gabor filter is applied to the left and right distorted images respectively, and natural statistical features un-der different scales and directions are extracted to act as monocular features. Then, left and right images are fused to construct a cyclopean map, and histograms of oriented gradient features are extracted from the cyclopean map to act as binocular features. Finally, a regression model between features and subjective scores is established via deep belief network. In the testing stage, based on the established regression model, left and right image quality scores are predicted and fused to get the final stereoscopic image quality score. Experimental results show that the proposed method is effective for both symmetrical and asymmetrical stereoscopic image databases, and can achieve high consistent alignment with subjective assessment.