结合JND模型的交叉验证深度图质量评价方法  

Depth assessment based on cross validation and JND model

作　　者：陈嘉丽[1] 彭宗举[1] 陈芬[1] 蒋刚毅[1] 郁梅[1] 

机构地区：[1]宁波大学信息科学与工程学院,浙江宁波315211

出　　处：《光电子．激光》2018年第1期85-94,共10页Journal of Optoelectronics·Laser

基　　金：国家"863"计划(2015AA015901);国家自然科学基金(61620106012;U1301257;61771269);浙江省自然科学基金(LY16F010002;LY15F010005;LY17F010005);宁波市自然科学基金(2015A610127;2015A610124);宁波大学科研基金(理)/学科(xkxl1502)资助项目

摘　　要：针对在3D视频(3DV)和自由视点视频(FVV)中传统的图像质量评价方法不适用于深度图的问题,本文从人类视觉感知特性出发,提出一种新的深度图质量评估算法。首先进行交叉验证,得到待评价深度图的差值图;然后提取遮挡掩膜,去除被遮挡的像素点;再根据人类视觉特性,考虑背景亮度掩蔽、纹理掩蔽和边缘敏感性等因素,应用恰可察觉失真(JND)模型得到每个像素点的误差可视阈值;最后计算错误像素率作为度量指标评价深度图的质量。实验结果表明,本文提出的算法能够准确地检测错误像素,所提出的度量指标与全参考度量指标的相关系数的平均值为0.833 0,最高达到0.933 8,与合成虚拟视点均方误差的相关系数的平均值为0.857 9,最高达到0.928 3。This paper presents a novel depth quality assessment scheme based on human visual perception characteristics. At first, cross validation is performed to get the difference map of the depth map to be assessed. Then the occlusion mask is extracted to remove the occluded pixels. According to the characteristics of human vision, considering the factors such as background luminance masking, texture masking and edge sensitivity,the error visibility threshold of each pixel is obtained by using the model of just noticea- ble distortion （JND）. Finally, the error pixels rate is calculated as a metric to assess the quality of the depth map. Experimental results show that the proposed algorithm can detect the error pixels accurately. The average value of the correlation coefficient between the proposed metric and the full reference metric is 0. 833 0 and the maximum is 0. 933 8. The average value of the correlation coefficient between the pro- posed metric and the mean square error of the synthesized virtual view is 0. 857 9 and the maximum is 0. 928 3.

关 键 词：深度图 质量评价 人类视觉特性 交叉验证 恰可察觉失真(JND)模型 

分 类 号：TN919.81[电子电信—通信与信息系统]