Super-Resolution Reconstruction of Depth Map Based on Non-Local Means Constraint

A super-resolution reconstruction method of depth map based on non-local means constraint is proposed to overcome the problems existing in most of the present super-resolution reconstructions,which are limited in local receptive fields with respect to the learning of mapping relationship and are based on the convolutional neural network.No up-sampling pre-processing of inputting low-resolution depth map is required in this method,and the low-resolution depth map will be up-sampled through sub-pixel assembly layer after training in the sub-pixel convolutional neural network.We make effective use of the non-local similarity of the depth map,and use a non-local means constraint on it to obtain the final reconstructed depth map,an end-to-end depth map super-resolution reconstruction network based on this is established.Experimental results on Middlebury dataset show that the algorithm in this paper achieves a marked improvement both on subjective visual effect and objective evaluation on RMSE and SSIM,compared with the existing RDN and EDSR algorithms,the reconstruction effect is superior.