Depth Image Enhancement Based On Low-cost Cameras

In recent years,with the continuous maturity of low-cost depth camera technology,depth images have been widely used in many computer vision fields,such as 3D reconstruction,indoor SLAM,virtual reality,augmented reality and other applications.However,low-cost cameras such as Kinect,Real Sense,etc.,usually have problems including low resolution,missing depth values,image holes,and noise.These problems will greatly affect the effect of technologies such as 3D reconstruction.So how to enhance the low-quality depth map is an urgent research topic.This paper firstly preprocesses the image captured by the camera,including camera calibration,distortion correction,RGB-D alignment,depth map hole filling and other operations.Make the quality of the depth map to achieve a good result at the original resolution.At the same time,in order to prepare for the subsequent depth map super-resolution recovery,the image alignment operation is performed on the color map and the depth map.This paper then proposes two color map-guided depth map super-resolution restoration models.One is a full convolution network model integrating the attention mechanism,and the other is a model integrating the technologies in three fields: single image super-resolution,guided super-resolution,and implicit representation of images.The latter uses the idea of guided super-resolution,borrows implicit representation of images as the skeleton,and use the module of super-resolution to code depth map.,thereby improving the effect of super-resolution restoration.At the same time,this paper compares with the current advanced methods on general data sets,and has achieved good results in many indicators.Finally,this paper shows the application demonstration of the model in the real scene,which proves the practicability of the method in this paper.