Hyperspectral images are useful in a variety of fields such as remote sensing, medical diagnosis, and agriculture. But it requires very expensive professional equipment and a lot of time to obtain. In this paper, we propose a deep learning architecture that reconstructs hyperspectral images from RGB images that are easy to acquire in real time. Hyperspectral reconstruction is inherently difficult because much information is lost when hyperspectral bands are integrated into three RGB channels. To effectively overcome the problem of hyperspectral restoration, we design a neural network with the following three basic principles. First, it adopts a method in which channels are gradually increased through several steps to restore hyperspectral images. Second, it is learned on a group basis for efficient restoration. Hyperspectral bands are divided into three groups: R, G, and B. Finally, the concept of channel back projection is newly proposed. In the process of gradually performing hyperspectral reconstruction, the reconstructed image is refined by repeatedly projecting the reconstructed hyperspectral to RGB. In the experimental results, these three principles proved the performance that exceeds the state-of-theart methods.