Frequency-Aware Degradation Modeling for Real-World Thermal Image Super-Resolution.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Super-resolution (SR) methods struggle with real-world thermal images due to simple degradation assumptions.
• Existing SR techniques lack generalization for complex thermal image characteristics.

Purpose of the Study:

• To develop an unsupervised SR framework for enhancing real-world thermal image resolution.
• To address the generalization limitations of supervised SR methods in thermal imaging.

Main Methods:

• Proposed a frequency-aware degradation model (TFADGAN) for thermal images.
• Employed adversarial learning with unpaired low-resolution (LR) thermal images to model degradation.
• Developed an SR model trained on pseudo-paired data generated by TFADGAN.

Main Results:

• TFADGAN effectively models complex degradation processes at different image frequencies.
• Generated degraded images closely resemble real-world LR thermal images in detail and contrast.
• The unsupervised SR framework improved PSNR by 1.28 dB and SSIM by 0.02 on real-world thermal images.

Conclusions:

• The proposed TFADGAN provides a reliable method for simulating LR thermal images.
• The unsupervised SR framework significantly enhances SR performance for real-world thermal sceneries.
• This approach offers a practical solution for improving thermal image resolution without paired data.