A Novel Domain Transfer-Based Approach for Unsupervised Thermal Image Super-Resolution

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Low-resolution thermal imaging presents challenges in detail and quality.
• Existing super-resolution methods struggle with thermal image domain limitations.

Purpose of the Study:

• To develop a transfer domain strategy for enhancing thermal image resolution.
• To address the limitations of low-resolution thermal sensors.

Main Methods:

• Utilized a CycleGAN architecture with a ResNet encoder and an attention module.
• Introduced a novel loss function for improved image generation.
• Trained the network on a multi-resolution thermal image dataset from three sensors.

Main Results:

• Achieved superior performance in thermal image super-resolution.
• Outperformed state-of-the-art methods in the 2nd CVPR-PBVS-2021 challenge.
• Generated higher-resolution thermal images of reasonable quality.

Conclusions:

• The proposed transfer domain strategy effectively enhances thermal image resolution.
• The method offers a viable solution for improving thermal imaging applications.
• The developed technique surpasses current benchmarks in thermal image super-resolution.