Thermal Image Guided Upsampling

Abstract

A low-resolution thermal imaging sensor becomes more affordable and is widely used in home applications. However, in order to understand detailed activities, a high-resolution thermal image is required. In this paper, we present an unsupervised deep learning framework for joint up-sampling which aims to generate higher resolution thermal image from given corresponding low-resolution thermal image and high-resolution RGB image. The proposed joint up-sampling framework is designed to learn linear transformation between high-resolution guided image and low-resolution thermal image. The new loss function is designed to minimize the error of the linear transformation and to preserve the edge information in the output high-resolution thermal image. Experimental results demonstrate that the proposed method produces a high-resolution thermal image which is comparable to expensive high-resolution thermal cameras.