Thermal colorization using deep neural network

Visual and thermal cameras have large modality gap while providing different information of the same scene. The spectral mapping of thermal imagery into color imagery is a challenging task due to inherit nonlinear relationship. This paper deals with the automatic colorization of thermal imagery into color images using deep encoder-decoder convolutional neural network architecture. The presented approach is trained and evaluated on an online thermal/color dataset, where the network is trained on the spectral mapping of thermal to color images. The training is automatic without requiring user intervention. Extensive experimentation is carried out on the online available dataset, demonstrating consistent and harmonious appearance of generated colorized images from input thermal images.