DuaDiff: Dual-Conditional Diffusion Model for Guided Thermal Image Super-Resolution

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Thermal imaging has low spatial resolution, limiting its applications.
• Enhancing thermal images with high-resolution visible images is challenging due to modality and resolution differences.

Purpose of the Study:

• To develop an innovative diffusion model for guided super-resolution (SR) of thermal images.
• To address the limitations of existing SR methods in handling significant modality and resolution gaps between thermal and visible images.

Main Methods:

• Introduced Dual-Conditional Diffusion (DuaDiff), a diffusion model with a dual-conditioning mechanism.
• Integrated a learnable Laplacian pyramid to extract multiscale high-frequency details from visible images.
• Utilized a semantic latent space projection and a multimodal latent feature cross-attention module for enhanced feature interaction.

Main Results:

• DuaDiff surpassed state-of-the-art methods in both visual quality and metric evaluations on FLIR-ADAS and CATS datasets for 4x and 8x SR.
• Demonstrated superior performance, particularly in scenarios with large resolution gaps between thermal and visible images.
• Confirmed DuaDiff's capability to recover high-fidelity semantic information in downstream tasks.

Conclusions:

• The proposed DuaDiff model effectively enhances thermal image super-resolution by leveraging complementary conditioning strategies.
• Combining Laplacian pyramid and semantic latent space conditioning provides robust performance across various resolution gaps.
• DuaDiff offers a promising solution for high-fidelity thermal image enhancement and semantic information recovery.