TherISuRNet - A Computationally Efficient Thermal Image Super-Resolution Network

Vishal Chudasama, Heena Patel, Kalpesh Prajapati, Kishor P. Upla, Raghavendra Ramachandra, Kiran Raja, Christoph Busch; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2020, pp. 86-87


Human perception is limited to perceive the objects beyond the range of visible wavelengths in the Electromagnetic (EM) spectrum. This prevents them to recognize objects in different conditions such as poor illumination or severe weather (e.g., under fog or smoke). The technological advancement in thermographic imaging enables the visualization of objects beyond visible range which enables it's use in many applications such as military, medical, agriculture, etc. However, due to the hardware constraints, the thermal cameras are limited with poor spatial resolution when compared to similar visible range RGB cameras. In this paper, we propose a Super-Resolution (SR) of thermal images using a deep neural network architecture which we refer to as TherISuRNet. We use progressive upscaling strategy with asymmetrical residual learning in the network which is computationally efficient for different upscaling factors such as x2, x3 and x4. The proposed architecture consists of different modules for low and high-frequency feature extraction along with upsampling blocks. The effectiveness of the proposed architecture in TherISuRNet is verified by evaluating it with different datasets. The obtained results indicate superior results as compared to other state-of-the-art SR methods.

Related Material

author = {Chudasama, Vishal and Patel, Heena and Prajapati, Kalpesh and Upla, Kishor P. and Ramachandra, Raghavendra and Raja, Kiran and Busch, Christoph},
title = {TherISuRNet - A Computationally Efficient Thermal Image Super-Resolution Network},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {June},
year = {2020}