Qualitative Colorization of Thermal Infrared Images using custom Convolutional Neural Networks

Abstract

The transformation of thermal infrared images to colored RGB images is a complex task. Thermal images with colorization present more information by virtue. This is based on the assumption that a human is more likely to notice features in a colorized image than in a grayscale thermal image. A fully automatic thermal infrared to visual color image transformation approach using custom trained Convolutional Neural Network (CNN) architecture is presented in the paper, in the absence of standard colorization methods for thermal images. The proposed model starts the training process from the ground up on a globally available OSU thermal color-pair dataset consisting of thermal images and corresponding colored images. The obtained results are quantitatively compared using standard evaluation metrics such as Peak signal-to-noise ratio (PSNR), Structural Similarity Index (SSIM), Root Mean Squared Error (RMSE), Feature Similarity Indexing Method (FSIM), and qualitatively evaluated by comparing perceptually realistic output images with ground truth.