Research On ToF Depth Map Denoising Based On Deep Neural Networks

In the era of information network and big data, people pay more attention to image data in-formation, different from the traditional two-dimensional image, the adding of depth infor-mation are creating a wide range of 3D visual applications. Based on the time of flight method, ToF camera can take the initiative to get the depth of the scene, and reflect the 3D information of scene and object’s surface accurately. By the interference of multiple error sources, original depth image obtained by the TOF camera contains a lot of noise, which affected its expansion and application.At present, the research scheme of ToF depth map denoising has been based on single frame image, and has not established a comprehensive noise model. In this paper, we mainly study a ToF depth image denoising algorithm based on deep convolutional neural network:1. This paper puts forward a complete deep convolutional neural network architecture with a mapping from image to image, the architecture uses a new Rectified linear unit and local re-sponse normalization techniques, extract and express the hidden feature of image blocks, and show good filtering effect on the natural image noise.2. The research uses the TOF camera and Kinect camera to build a TOF depth image ground truth dataset. Through camera calibration and depth map registration, we use Kinect fu-sion algorithm to capture the true depth of scene. And combined with the exposure properties of the TOF camera, we create the TOF depth map ground truth dataset.3. A ToF depth image denoising algorithm based on depth network is proposed. By training the TOF depth map ground truth dataset, the convolutional neural network learns the character-istics of ToF’s noise implicitly. Experiments show that the algorithm can eliminate the original error of ToF depth map effectively, and keep the depth details of the edge and boundary region of the object.