Multi-Source Deep Residual Fusion Network for Depth Image Super-resolution
Xiaohui Hao, Tao Lü, Yanduo Zhang, Zhongyuan Wang, Hui Chen · 2019
Comparing with color images, depth images are often in lack of texture information in high quality. Depth image super-resolution provides an efficient solution to enhance the high frequency information of LR depth image. In this paper, we propose a novel multi-source residual fusion neural network named "MSRFN", which fully uses the fruitful texture information of color images to guide the depth images reconstruction. Initially, color and depth images are used to extract residual features in two-branch network. Then, color residual and depth residual are fused by the fusion network. Finally, the high-resolution (HR) depth map is reconstructed by fusing multi-source high-frequency information. Experimental results on MPI Sintel and Middlebury public databases show that MSRFN outperforms some state-of-the-art approaches in subjective and objective measures.