Cross-Domain Feature Extraction Using CycleGAN for Large FoV Thermal Image Creation

Thermal images are extensively utilized in defense operations due to their ability to capture thermal radiation. However, the limited Field of View (FoV) in thermal imaging systems often results in insufficient and fragmented large Field of View (FoV) simulation data, thereby compromising the effectiveness of defense applications using large Field of View (FoV) cameras. To address this challenge, we propose an innovative approach employing image stitching techniques to create comprehensive and detailed views. Traditional image feature extractors such as Scale-Invariant Feature Transform (SIFT) and Speeded Up Robust Features (SURF) are optimized for Red, Green, and Blue (RGB) images. Hence, their performance on thermal images is often sub-optimal as the thermal images contain low-resolution, reduced contrast and dynamic range, non-uniform thermal emissions, and the absence of color information, making identifying and matching robust features and descriptors difficult. To enhance feature extraction, we adopt the Cycle Generative Adversarial Network (CycleGAN) technique to convert thermal images into their RGB counterparts prior to feature extraction. This enables us to leverage the robustness of RGB-optimized feature extractors and amalgamate them with the features obtained from the original thermal images, resulting in a richer and more comprehensive feature set. The performance of this proposed network has been demonstrated and validated on various datasets, showing promising results for large Field of View (FoV) thermal image creation.