加权SIFT流深度迁移的单幅图像2D转3D  

Single Image 2D-to-3D Conversion via Weighted SIFT Flow

作　　者：袁红星[1] 吴少群[1] 朱仁祥[1] 安鹏[1] 

机构地区：[1]宁波工程学院电子与信息工程学院,浙江宁波315016

出　　处：《电子学报》2015年第2期242-247,共6页Acta Electronica Sinica

基　　金：国家自然科学基金(No.61071173);浙江省自然科学基金(No.LY12F01001;No.Y1100253);宁波市自然科学基金(No.2012A610043;No.2012A610111;No.2011A610186);浙江省教育厅科研项目(No.Y201431834)

摘　　要：2D视频转3D视频是解决3D片源不足的主要手段,而单幅图像的深度估计是其中的关键步骤.提出基于加权SIFT流深度迁移和能量模型优化的单幅图像深度提取方法.首先利用图像的全局描述符从深度图数据库中检索出近邻图像;其次通过SIFT流建立输入图像和近邻图像之间像素级稠密对应关系;再次由SIFT流误差计算迁移权重,将近邻图像对应像素点的深度乘以权重后迁移到输入图像上;然后利用均值滤波对迁移后的近邻图像深度进行融合;最后建立深度图优化能量模型,在尽量接近迁移后近邻图像深度的前提下,平滑梯度较小区域的深度.实验结果表明,该方法降低了估计深度图的平均相对误差,增强了深度图的均匀性.2D-to-3D conversion is one of the important ways to alleviate the lack of 3D content,in which the key technique is depth estimation from a single image. A depth extraction method based on weighted SIFT flowdepth transferring and energy model optimization is proposed. First,k-nearest neighbor images were retrieved by evaluating the global image descriptors between the target image and the images from the RGBD database. Then,pixels of the target image and pixels of its neighbors in the database were matched using SIFT flow,which also generated matching errors used to determine the transferring weights of the neighbors. Next,depth maps of neighboring images were transferred to the target image according to the pixel matching and the transferring weights. The final depth map was obtained by depth refinement based on an energy model. Experimental results showthat our scheme can significantly reduce the average relative error and improve the uniformity of the depth map.

关 键 词：2D转3D 尺度不变特征变换流 深度估计 深度图优化 能量模型 

分 类 号：TN911.73[电子电信—通信与信息系统]