结合相机阵列选择性光场重聚焦的显著性检测  

Saliency detection combined with selective light field refocusing of camera array

作　　者：冯洁 王世刚[1] 韦健[1] 赵岩[1] FENG Jie;WANG Shi-gang;WEI Jian;ZHAO Yan(College of Communications Engineering,Jilin University,Changchun 130012,China)

机构地区：[1]吉林大学通信工程学院,吉林长春130012

出　　处：《中国光学》2021年第3期587-595,共9页Chinese Optics

基　　金：国家自然基金重点项目(No.61631009);国家十三五重点研发计划项目(No.2017YFB0404800);中央高校基本科研业务费专项资金(No.2017TD-19)。

摘　　要：针对现有方法处理包含多个显著目标以及显著目标的某些区域与背景区域对比不明显的场景所得显著图不够精细,甚至会丢失某些显著性区域的不足,本文提出了一种结合相机阵列选择性光场重聚焦的显著性检测方法。选用光场数据集,利用同一场景的多幅视点图像,首先对中心视点图像进行结合超分辨率的重聚焦渲染;然后利用基于图的显著性检测方法提出结合全局和局部平滑度约束的传播模型以防止错误标签传播,得到的显著性粗图经过目标图的细化后最终输出精细的检测结果。另外,对于包含多个显著目标的场景,通过选择对场景中某一深度层进行重聚焦,同时对其他深度层产生不同程度的模糊,可以更精确、细致地检测出位于该深度层上的显著目标,一定程度上实现了可选择的显著性检测。在4D光场数据集上进行了实验,结果表明:本文提出的方法所得显著图与真值图之间的平均绝对误差的均值为0.212 8,较现有方法有所降低,检测结果包含更丰富的显著性目标信息,改善了现有显著性检测方法的不足。For the multiple salient targets scene,as well as a scene in which some areas of the salient target do not contrast significantly with the background area,the saliency maps obtained by existing algorithms are not fine enough or even lose some saliency regions.In this paper,a new saliency detection method combined with selective light field refocusing of camera array is proposed.In this method,the light field dataset is selected and multi-viewpoint images of the same scene are used.First,we perform refocusing rendering combined with super-resolution on the central viewpoint image.Then,on the basis of the graph-based saliency detection method,we propose a propagation model combining global and local smoothness constraints to prevent false label propagation.Finally,the obtained coarse saliency map is refined through the object map to output the final saliency map.In addition,for the scene that contains multiple salient targets,by refocusing a certain depth layer in the scene,and producing varying degrees of blurring to other depth layers,the salient targets on the depth layer can be detected accurately and in detail.To a certain extent,the optional saliency detection is realized.Experiments on the 4D light field dataset show that the average Mean Absolute Error(MAE)between the ground truth and the saliency map obtained by the method proposed in this paper is 0.2128,which is lower than those obtained by the existing methods.The detection result contains more detailed information about the salient target,which improves the above-mentioned shortcomings of the existing salient detection methods.

关 键 词：相机阵列 重聚焦 显著性检测 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]