Noise Reduction of Depth Cameras Images Based on Deep Neural Network

Today, infrared sensors or depth sensors are widely used to control applications, games, information acquisition, dynamic and static 3D scenes. Despite the widespread use of these images, their quality is limited to low-quality images, as the infrared sensor does not have high resolution and the images produced by it have noise. Therefore, given the problems and the importance of using 3-D images, the quality of these images should be improved in order to provide accurate images from depth cameras. In this paper, the noise reduction of depth images using convolutional neural networks is considered. A convolutional neural network with a depth of 20 and three layers and a pre-trained neural network is used. We developed the system and tested its performance for two datasets of depth and color images, Middlebury and EURECOM Kinect Face. Results show that for EURECOM Kinect Face images, PSNR improvement is approximately 8 to 15 dB and for Middlebury images the PSNR improvement is about 5 to 14 dB.