This study propose a robust 3D depth-map generation algorithm using a single image. Unlike previous related works estimating global depth-map using deep neural networks, this study uses the global and local feature of image together to reflect local changes in the depth map instead of using only global feature. A coarse-scale network is designed to predict the global-coarse depth map structure using a global view of the scene and the finer-scale random forest (RF) is to be designed to refine the depth map based on combination of original image and coarse depth map. As the first step, we use a partial structure of the multi-scale deep network (MSDN) to predict the depth of the scene at a global level. As the second step, we propose local patchbased deep RF to estimate the local depth and smoothen noise of local depth map by combining MSDN global-coarse network. The proposed algorithm was successfully applied to various single images and yielded a more accurate depthmap estimation performance than other existing methods.