Near-Infrared Image Colorization by Convolutional Neural Network with Perceptual Loss
Shoichiro Sekiguchi, Masao Yamamoto · 2020
Near-infrared (NIR) imaging is widely used in various fields, such as surveillance and autonomous driving, because the NIR frequency band is invisible to the human eye. NIR images are usually represented in grayscale, which may disagree with human cognition. Thus, an RGB-color-model representation is preferable for human perception. We proposed a deep neural network for NIR image colorization that employs an encoder-decoder-based convolutional neural network for the colorization. For a realistic colorization, we used perceptual loss and pixelwise loss functions. Experimental results reveal that the proposed network can be suitably generalized to unseen data owing to perceptual loss.