基于超像素分割的深度图像修复算法  

Depth map inpainting algorithm based on superpixel segmentation

作　　者：胡天佑[1] 彭宗举[1] 焦任直 陈芬[1] 左力文[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》2016年第10期1120-1128,共9页Journal of Optoelectronics·Laser

基　　金：国家自然科学基金(U1301257;61271270);国家"863"计划(2015AA015901);浙江省自然科学基金(LY16F010002;LY15F010005);宁波市自然科学基金(2015A610127;2015A610124)资助项目

摘　　要：为了提高Kinect相机获取的深度图质量,提出了一种基于超像素分割的图像修复算法。首先对深度图和彩色图分别进行双边滤波和超像素分割;其次,结合深度图像和彩色图像相似性,记录彩色分割块的位置,并对应于深度图中;最后,在每个分割块对应深度区域中,根据丢失像素点在分割块中所占比例,划分为无空洞区域、小空洞区域、大空洞区域和全空洞区域4类。采用快速行进算法对小空洞区域进行修复,利用中值填补算法进行大空洞区域修复,对全空洞区域利用邻域区间对应彩色图像相似性进行填充。4种类型中的无空洞区域无需修复。实验结果表明,本文方法与FMM、Shen和Scheming的方法相比,平均均方根误差(RMSE)分别降低了2.958 4、0.822 9和0.078 0,修复的主观质量也有所提高。The accuracy of depth scene from Kinect will always be influenced by many surrounding factors.Therefore missing areas will appear in the depth map obtained by depth camera.This paper proposes an inpainting algorithm based on superpixel segmentation to improve the quality of depth map.Firstly,the depth map and color map will be filtered and segmented,respectively.Secondly,the similarity of color image and depth map will be used for K-means cluster,and each cluster position is recorded as labels which will be mapped to the depth map.Finally,clusters are divided into four types according to the ratio of the missing pixels in each cluster.The types include cluster without holes,cluster with small holes,cluster with large holes and all-hole cluster.The clusters with small holes and large holes will be filled with fast marching method and median filling algorithm,respectively.Besides,we use nearly similar area filling method for the all-hole cluster,and the processing of cluster without holes is unnecessary.Experimental results show that the residual mean square error of the proposed method is lower than that of FMM,Shen′s and Scheming′s methods by 2.958 4,0.822 49 and 0.078 40,respectively.Besides,the subjective quality of the proposed method is better than that of those methods.

关 键 词：深度图 深度图空洞修复 超像素分割 KINECT 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]