基于语义分割的MRF深度图修复

MRF Depth Map Restoration with Semantic Segmentation

由于自身成像原理的限制和自然环境的影响,通过深度相机获取的深度图存在深度缺失和测量噪声。针对上述两种问题,提出了语义分割先验的深度缺失处理和深度图平滑降噪方法。首先,彩色图像输入语义分割网络进行分割;然后,利用深度数据和彩色数据的空间、结构相关性,结合分割结果,填充深度空洞;最后,利用马尔可夫随机场,结合分割结果,对深度图进行修复。实验结果表明,该算法对不同的深度空洞鲁棒性更好,修复后的图像PSNR值至少提高了7.7%。

The depth map obtained by the depth camera has depth missing and measurement noise due to the limitations of its own imaging technique and the influence of the natural environment.For the aforementioned two issues,the semantic segmentation a priori methods of depth map smoothing and noise reduction are suggested.The depth voids are first filled using the spatial and structural correlation between the depth data and the color data in conjunction with the segmentation findings,and then the depth map is repaired using Markov random field in conjunction with the segmentation results.The experimental results demonstrate that the algorithm in this study is more resistant to various depth voids,and the restored image's PSNR is enhanced by at least 7.7%.