Depth Image Noise Reduction and Super-Resolution by Pixel-Wise Multi-Frame Fusion

High-quality depth images are required for stable and accurate computer vision. Depth captured by cameras tend to be noisy, incomplete, of low-resolution. Therefore, increasing the accuracy resolution is desirable. We propose a method reducing noise holes from pixel pixel, resolution. For each in target image, linear space focal point camera through existing object divided into equally spaced grids. In grid, difference grid surface obtained multiple tracked images, which have noisy values respective image pixels. Then, coordinates correct obtainable random noise. The missing completed. can also increased creating new pixels between then using same process as that used reduction. Evaluation results demonstrated proposed do processing with less GPU memory. Furthermore, was able reduce more accurately, especially around edges, details objects than conventional method. super-resolution produced high-resolution smoother edges methods.