Laplacian reconstructive network for guided thermal super-resolution.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Multi-modal data fusion is crucial for applications like robotics and autonomous navigation.
• High-resolution visual data is often limited by sensor cost and quality.
• Existing fusion methods require dense models and heavy computation.

Purpose of the Study:

• To introduce LapGSR, a lightweight, multimodal generative model for guided thermal super-resolution.
• To overcome the computational burden of dense models in image fusion.
• To improve image resolution using RGB and thermal data fusion.

Main Methods:

• Utilized Laplacian image pyramids on RGB data to extract edge information.
• Employed a lightweight generative architecture for efficient feature extraction.
• Combined pixel and adversarial loss functions for enhanced super-resolution.
• Developed a multimodal approach fusing RGB and thermal imagery.

Main Results:

• LapGSR achieved high-resolution thermal images while preserving spatial and structural details.
• The model demonstrated significant efficiency with fewer parameters compared to state-of-the-art (SOTA) models.
• Excellent performance was observed on the ULB17-VT and VGTSR cross-domain datasets.

Conclusions:

• LapGSR offers an efficient and compact solution for thermal super-resolution.
• The proposed method effectively fuses multi-modal data for improved image quality.
• LapGSR presents a viable alternative to computationally intensive fusion models.