基于不舒适感知与组稀疏编码的立体图像视觉舒适度评价

Stereoscopic Image Visual Comfort Prediction via Discomfort-aware Group Sparse Coding

针对现有的需要依靠映射模型来缩小图像特征和视觉舒适度之间的差距的立体图像视觉舒适度评价方法,提出了一种基于不舒适感知与组稀疏编码的立体图像视觉舒适度评价方法。首先,考虑人眼的视觉注意机制和双目融合极限,结合显著不舒适区域和空间频率提取视差统计特征;然后,对立体图像特征按照主观评分值(MOS)进行不舒适程度分组,并利用组聚类构建不舒适感知字典;最后,通过加权在字典上进行组稀疏编码得到组稀疏系数和设定的每组不舒适程度值,预测得到立体图像视觉舒适度的客观评价值。实验结果表明,对于IVY LAB立体图像库,所提出方法的Pearson线性相关系数(PLCC)值在0.89以上,Spearman等级相关系数(SROCC)值在0.87以上,优于相关代表性的立体图像视觉舒适度评价方法,表明所提出方法的客观评价值与主观评价值具有良好的一致性,更加符合人眼视觉系统的感知特性。

According to the mapping model which needs to bridge the gap between the image features and visual comfort in current stereoscopic image visual comfort prediction, a new stereoscopic image visual comfort assessment method via discomfort-aware group sparse coding is proposed. Firstly, the disparity statistics are extracted combined with salient discomfort region and spatial frequency by considering the visual attention mechanism and binocular fusion limit. Then, a discomfort-aware dictionary is trained covering different levels of discomfort degree by using clustering. Finally, the objective scores of stereoscopic image visual comfort is predicted by weighting the group sparse coefficients obtained by group sparse coding in dictionary. The proposed method is tested on the IVY LAB stereoscopic image database, and the experimental results show that for the proposed method, the Pearson linear correlation coefficient reaches 0. 89, and the Spearman rank correlation coefficient reaches 0. 87, significantly better than state-of-the-art stereoscopic image visual comfort assessment methods. It is indicated that the proposed method can achieve higher correlation with subjective scores, and accords with the human visual perception.