Depth Map Enhancement Under The Guidance Of Color Image

As a way to represent3D information, depth maps have been widely used in computer vision systems. However, depth maps captured by different sensors, such as lidar, TOF camera, Kinect, are not perfect. They may suffer from defects such as low resolution, unknown regions, and noise. This thesis researches on how to get a high-quality depth map from a defective depth map based on digital image processing algorithms. This task is referred to as the depth map enhancement.In this thesis, algorithms combine the aligned color image to do depth enhancement. They can be divided into two parts, local-based methods and global-based methods.This thesis first proposes a hole-filling algorithm via a fast marching method in depth maps. This method incorporates the guidance color image into the traditional inpainting method based on fast marching method. It uses a new depth estimation model and modifies the order to do the depth inpainting. The algorithm solves the problem that depth maps captured by Kinect have unknown regions, and outperforms other methods in large unknown regions.In addition, this thesis proposes a depth enhancement method via anisotropic diffusion. The pixels with known depth values are treated as the heat sources and the depth enhancement is performed via diffusing the depth from these sources to unknown regions. The diffusion conductivity is designed in terms of the guidance color image. The proposed algorithm is competent to other method when evaluated in different datasets. What’s more, the thesis incorporates the window-based data constraints and second-order smoothness into the original Markov random field model and shows the framework of representation and inference.