基于RGB引导的深度图增强方法

RGB-guided depth image ehancement method

受ToF相机成像原理和硬件设备的限制,深度图分辨率较低,且包含较多噪声与空洞。该问题的存在对位姿估计、尺寸测量等物料识别与测量任务带来较大影响。针对上述问题,文中利用RGB图像与深度图之间的相关性,使用高质量RGB图像引导深度图的降噪、补洞与超分辨。通过深度神经网络融合RGB与深度图像信息,并结合部分卷积网络对深度图空洞补全,重建深度图质量有显著提升。

Due to the limitations of ToF camera,including image-forming principle and hardware equipment,the depth image has a low resolution and contains noise and holes.This problem has a great impact on pose estimation,size measurement and other material identification,and measurement tasks.In this paper,we propose a neural network for denoising,hole filling and super-resolution simultaneously guided by RGB image,which is based on correlation between RGB images and depth maps.Through the fusion of convolutional neural network with RGBD information,and combined with the use of partial convolution to fill hole,the quality of reconstructed depth map is significantly improved.