A feature refinement and adaptive generative adversarial network for thermal infrared image colorization

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Colorizing thermal infrared images is challenging due to unrealistic saturation and texture.
• Existing methods often fail to preserve detailed and semantic information.

Purpose of the Study:

• To develop an advanced Generative Adversarial Network for high-quality thermal infrared image colorization.
• To address limitations in detail, semantics, and texture in current colorization techniques.

Main Methods:

• Proposed the Feature Refinement and Adaptive Generative Adversarial Network (FRAGAN).
• Introduced Residual Feature Refinement Module (RFRM) for accuracy and generalization.
• Utilized Feature Adaptation Module (FAM) to prevent sub-region information loss.
• Incorporated Trinity Attention Module (TAM) for capturing local semantic features.

Main Results:

• FRAGAN demonstrated superior performance metrics and visual quality compared to state-of-the-art methods.
• Colorized images exhibited enhanced clarity, realism, and preserved texture.
• Experiments conducted on KAIST and FLIR datasets validated the proposed approach.

Conclusions:

• FRAGAN effectively enhances detailed, semantic, and contextual information for thermal image colorization.
• The proposed modules significantly improve the accuracy and generalization of the colorization model.
• FRAGAN represents a significant advancement in realistic thermal infrared image colorization.