Infer Thermal Information from Visual Information: A Cross Imaging Modality Edge Learning (CIMEL) Framework.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Infrared (IR) thermography suffers from low spatial resolution, limiting measurement accuracy compared to digital cameras.
• High spatial resolution in IR imaging is often cost-prohibitive or technically infeasible due to high sample rate requirements.
• Improving IR image quality, especially edge definition, is critical for applications like surveillance and industrial inspection without hardware upgrades.

Purpose of the Study:

• To develop a novel framework for enhancing edge details in infrared images.
• To leverage corresponding visual images for learning high-frequency features to improve IR image quality.
• To enable effective IR image enhancement without requiring new or upgraded hardware.

Main Methods:

• A Conditional Generative Adversarial Network (CGAN)-based framework was proposed.
• A U-Net generator was employed for cross-modality learning.
• A dual-discriminator architecture was introduced, focusing separately on edge and content/background features.
• The framework learns high-frequency features from visual images to enhance low-resolution IR images.

Main Results:

• The proposed framework successfully enhanced barely visible edges in IR images.
• The enhancement process preserved the original content information without introducing artifacts.
• The method demonstrated effective edge enhancement in low-resolution IR images.
• The dual-discriminator guided the generator effectively in learning high and low-frequency features.

Conclusions:

• The CGAN-based framework offers an effective solution for enhancing IR image edges.
• This approach improves the quality of IR images without the need for hardware upgrades.
• The method's ability to work with only IR images during testing enhances its applicability in various scenarios, including active thermography.