Article: Deep learning-based lightweight approach to thermal super resolution Journal: International Journal of Biometrics (IJBM) 2023 Vol.15 No.3/4 pp.505 - 520 Abstract: In this paper, we propose a thermal image super-resolution (SR) technique using a lightweight deep learning model which we refer to as thermal lightweight network (TherLiNet). We refine interpolated images using convolutional layers interleaved with different activation functions along with residual learning in the network. The effectiveness of the proposed architecture is evaluated against widely used deep learning-based super resolution models namely, super-resolution convolutional neural network (SRCNN), thermal enhancement network (TEN) and very deep super resolution (VDSR). Training and testing is done with different thermal datasets using different scale factors. To further explore the possibilities, red green blue (RGB) guided training is also performed and evaluated on the thermal image datasets. Peak signal to noise ratio (PSNR) and structural similarity index (SSIM), the most widely accepted parameters have been used for evaluation of the proposed model. The model is also compared to other models based on computation time to generate results. We also demonstrate the results in terms of qualitative values of the model compared to other super-resolution (SR) techniques. Inderscience Publishers - linking academia, business and industry through research

Title: Deep learning-based lightweight approach to thermal super resolution

Authors: Shashwat Pandey; Darshika Sharma; Basant Kumar; Himanshu Singh

Addresses: Computer Science and Engineering Department, MNNIT Allahabad, Prayagraj, India ' Electronics and Communication Department, MNNIT Allahabad, Prayagraj, India ' Electronics and Communication Department, MNNIT Allahabad, Prayagraj, India ' IRDE, DRDO, Dehradun, India

Abstract: In this paper, we propose a thermal image super-resolution (SR) technique using a lightweight deep learning model which we refer to as thermal lightweight network (TherLiNet). We refine interpolated images using convolutional layers interleaved with different activation functions along with residual learning in the network. The effectiveness of the proposed architecture is evaluated against widely used deep learning-based super resolution models namely, super-resolution convolutional neural network (SRCNN), thermal enhancement network (TEN) and very deep super resolution (VDSR). Training and testing is done with different thermal datasets using different scale factors. To further explore the possibilities, red green blue (RGB) guided training is also performed and evaluated on the thermal image datasets. Peak signal to noise ratio (PSNR) and structural similarity index (SSIM), the most widely accepted parameters have been used for evaluation of the proposed model. The model is also compared to other models based on computation time to generate results. We also demonstrate the results in terms of qualitative values of the model compared to other super-resolution (SR) techniques.

Keywords: thermal images; super resolution; deep learning; RGB guidance; thermal lightweight network; TherLiNet; thermal super resolution.

DOI: 10.1504/IJBM.2023.130650

International Journal of Biometrics, 2023 Vol.15 No.3/4, pp.505 - 520

Received: 17 Jul 2021
Accepted: 29 Mar 2022
Published online: 02 May 2023 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article

Title: Deep learning-based lightweight approach to thermal super resolution

Keep up-to-date