基于权重组合学习的无参考立体图像质量评价

深度学习被广泛应用于2D 图像的质量评价(2D-IQA) 研究中,而在3D 图像质量评价(3D-IQA) 中还没有展开深入研究.针对对齐失真立体图像,本文提出了一种基于权重组合学习的无参考立体图像质量评价模型.通过有效融合两支独立的单视图质量评价深度网络模型,将左右眼视图作为整体对象进行评估; 再根据双目竞争原理,又设计了一种权重深度网络用以估计左右眼的不同能量分布; 最后,这两个子网络组合成端到端的权重组合学习深度质量网络.实验结果证明: 该模型对于对称失真的立体图像质量评价有显著提升.

Blind Symmetrically Distorted Stereoscopic Images Quality Assessment Based on Weighted Ensemble Learning

Recently deep learning has been largely applied to 2D image quality assessment(2D-IQA) but rarely to 3D image quality assessment(3D-IQA). In this letter, we propose a new method for blind symmetrically distorted stereoscopic images quality assessment utilizing multiple features fusion in deep network to evaluate the left-and-right views as an integration with no extra cost. According to binocular rivalry, a weighted ensemble learning network is developed for learning energy of dominant eye. We integrate these two networks into a full end-to-end network called a Weighted Ensemble Deep Quality Network(WEDQN). Our experimental results can demonstrate that the proposed method leads to significant improved quality prediction of symmetrically distorted stereoscopic images.