多核学习纹理特征的立体图像质量评价

为了有效地评价各种失真类型双目立体图像的质量,提出利用多核学习机学习立体图像平面纹理信息和3D映射信息的通用无参考立体图像质量评价IQA方法。该方法首先利用立体匹配模型对左右视图进行处理,获得相应的视差图DM和误差能量图DMEE;对左右视图、视差图和误差能量图进行相位一致性和结构张量变换,获得它们的平坦区和边缘区;分别提取左右视图两个区域纹理特征作为平面信息,提取视差图的纹理特征和误差能量图的统计特征作为3D信息;将所有特征作为多核学习机的输入,利用多核学习的信息融合能力预测待测失真立体图像质量。由于充分利用了立体图像的左右视图、视差图和误差能量图的失真信息,以及多核学习的信息融合能力,该方法具有很好的前景。在LIVE 3D图像质量数据库上的实验表明,该方法与主观质量有较高一致性,与现有的双目立体质量评价方法相比有很大的竞争力。

Quality assessment of stereoscopic images with texture features via multiple kernel learning

To assess various distorted stereoscopic image quality efficiently, we propose a no reference stereoscopic image quality assessment model that utilizes 3D map information and 2D texture information to drive the multiple kernel learning machine (MKL). Firstly, the model utilizes the stereoscopic matching model to obtain disparity map and disparity map of error energy on the basis of left view and right view. Secondly, left view, right view, disparity map and disparity map of error energy are all transformed by phase congruency and structure tensor to obtain their marginal zone and planar zone. Thirdly, the model extracts the texture feature of the two zones from left view and right view respectively as plane information, and it extracts statistics feature and texture feature of the two zones respectively from disparity map and disparity map of error energy as 3D information. Finally, all features are input to the MKL to predict the quality of tested images. Experiments on the LIVE 3D image quality database demonstrate that the proposed method has good consistency with human subject quality and has high competitiveness in comparison with the state-of-the-art models.