基于深度图预处理和图像修复的虚拟视点绘制

Virtual View Rendering Based on Depth Map Preprocessing and Image Inpainting

基于深度图像的绘制(DIBR)技术是绘制虚拟视点图像的重要方法.针对单视点虚拟视图绘制过程中的空洞、裂缝、重叠问题,提出一种基于深度图预处理和图像修复的绘制算法.首先根据虚拟视点的变换方向对深度图进行预处理,减缓前景到背景的深度突变,将大面积空洞分割为多个较小的空洞;然后结合一投四算法和Z-buffer算法对DIBR过程进行改进,改善裂缝和像素点间的重叠;最后使用深度信息引导图像修复算法,使空洞修复从背景一侧开始且优先选择背景纹理.在Matlab环境下,使用微软研究院提供的3D图像序列进行虚拟视点绘制并比较与真实图像的相似度,实验结果表明,该算法的平均峰值信噪比为27.4747dB,比Criminisi算法和递进填充线算法分别高出5.65%和2.97%;结构相似度为0.7708,比2种已有算法分别高出1.22%和0.80%.

Depth-image-based-rendering(DIBR)is an effective method for virtual view synthesis.To solve the problems in single view rendering such as hole,crack and pixel overlap,a virtual view rendering method based on depth map preprocessing and image inpainting is proposed.Firstly,the depth map is preprocessed based on the transformation direction.The sharp depth transition from foreground to background is mitigated and disocclusion regions are divided into several small holes.Secondly,the DIBR process is modified by combining the one-to-four algorithm and Z-buffer algorithm.Finally,the depth information is used to guide the image inpainting algorithm,so that the hole filling starts from the background side and the background texture is preferred.In the Matlab environment,the 3D image sequences provided by Microsoft Research are used to generate the virtual view,and the algorithm is evaluated by comparing the similarity between the virtual view and the real target view.Experimental results show that the average PSNR of proposed method is 27.4747dB,which is 5.65%higher than Criminisi algorithm and 2.97%higher than recursive line-based algorithm.The average SSIM is 0.7708,1.22%and 0.80%higher than the two existing algorithms,respectively.