Infrared image colorization using an edge aware auto encoder decoder with the multi-resolution fusion

摘要

This paper propoes a novel method for near infrared (NIR) image colorization. The proposed method have two parts. In the first part, the RGB output is generated by the edge aware auto encoder decoder, which includes symmetrical jump connections for precise positioning and assistant training networks for enhancing edge and color regions. In the second part, the details of the input NIR image are integrated with the original RGB output, which is a multi-resolution fusion guided by weight maps. No user interaction and similar image is required in our method but also retains clear edges and rich details. We have used a large number of NIR image datasets to evaluate the method in our paper. In the experiments, we show in detail that the proposed method works well and outperforms contemporary state-of-the-art approaches.