Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Thermal Image Calibration and Correction using Unpaired Cycle-Consistent Adversarial Networks

Hossein Rajoli, Pouya Afshin, Fatemeh Afghah

TL;DR

This study tackles the scarcity and heterogeneity of aerial wildfire datasets by proposing a CycleGAN-based calibration framework that translates low-quality IR images to a high-quality, temperature-referenced domain. The model employs non-symmetric generators with multi-level feature fusion and RGB attribute conditioning to preserve content while enhancing fidelity, guided by a composite loss that includes adversarial, cycle, identity, perceptual, and SSIM components. Evaluation on Flame II and Flame III datasets shows improved structural similarity and perceptual fidelity over baselines, indicating better alignment with high-tech camera outputs. The approach enables harmonization of disparate wildfire imagery, facilitating scalable DL-based wildfire analysis across differing sensor platforms and improving data utility for detection, characterization, and management tasks.

Abstract

Unmanned aerial vehicles (UAVs) offer a flexible and cost-effective solution for wildfire monitoring. However, their widespread deployment during wildfires has been hindered by a lack of operational guidelines and concerns about potential interference with aircraft systems. Consequently, the progress in developing deep-learning models for wildfire detection and characterization using aerial images is constrained by the limited availability, size, and quality of existing datasets. This paper introduces a solution aimed at enhancing the quality of current aerial wildfire datasets to align with advancements in camera technology. The proposed approach offers a solution to create a comprehensive, standardized large-scale image dataset. This paper presents a pipeline based on CycleGAN to enhance wildfire datasets and a novel fusion method that integrates paired RGB images as attribute conditioning in the generators of both directions, improving the accuracy of the generated images.

Thermal Image Calibration and Correction using Unpaired Cycle-Consistent Adversarial Networks

TL;DR

This study tackles the scarcity and heterogeneity of aerial wildfire datasets by proposing a CycleGAN-based calibration framework that translates low-quality IR images to a high-quality, temperature-referenced domain. The model employs non-symmetric generators with multi-level feature fusion and RGB attribute conditioning to preserve content while enhancing fidelity, guided by a composite loss that includes adversarial, cycle, identity, perceptual, and SSIM components. Evaluation on Flame II and Flame III datasets shows improved structural similarity and perceptual fidelity over baselines, indicating better alignment with high-tech camera outputs. The approach enables harmonization of disparate wildfire imagery, facilitating scalable DL-based wildfire analysis across differing sensor platforms and improving data utility for detection, characterization, and management tasks.

Abstract

Unmanned aerial vehicles (UAVs) offer a flexible and cost-effective solution for wildfire monitoring. However, their widespread deployment during wildfires has been hindered by a lack of operational guidelines and concerns about potential interference with aircraft systems. Consequently, the progress in developing deep-learning models for wildfire detection and characterization using aerial images is constrained by the limited availability, size, and quality of existing datasets. This paper introduces a solution aimed at enhancing the quality of current aerial wildfire datasets to align with advancements in camera technology. The proposed approach offers a solution to create a comprehensive, standardized large-scale image dataset. This paper presents a pipeline based on CycleGAN to enhance wildfire datasets and a novel fusion method that integrates paired RGB images as attribute conditioning in the generators of both directions, improving the accuracy of the generated images.
Paper Structure (11 sections, 1 equation, 4 figures, 1 table)

This paper contains 11 sections, 1 equation, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Works
  3. Proposed Method
  4. Proposed Architecture
  5. Non-Symmetric Generator Architecture
  6. Discriminator Architecture
  7. Loss Functions
  8. Evaluation
  9. Dataset
  10. Implementation and Results
  11. Conclusion

Figures (4)

  • Figure 1: The schematic diagram of the proposed pipeline
  • Figure 2: The schematic diagram of both proposed generators
  • Figure 3: Flame 2 samples. Due to lack of temperature reference, cold areas are demonstrated with red color.
  • Figure 4: Qualitative results from 50 epochs comparing the proposed IR-to-IR image translation method against other translation models. The first two Columns depict the input RGB and thermal images, followed by outputs from Cycle-GAN, U-Net, and our method, both with and without the super-resolution upsampling module.