Resilient recognition system for degraded thermal images using convolutional neural networks

For biometric identity applications, thermal infrared face recognition technologies have become a powerful alternative to visual systems. However, thermal images can undergo degradation in various ways, including noise, blurring, reduced spatial resolution, and temperature drift, in addition to being affected by changes in pose and facial expression. In this paper, we propose using convolutional neural networks (CNNs) to recognise degraded thermal face images. The system deals efficiently with poor-quality images resulting from various causes. We describe how a CNN structure processes images and use ResNet-50 architecture to demonstrate our results, being an essential deep learning model that has proven its efficiency in various computer vision and machine learning applications. We conduct experiments under different conditions and scenarios that tackle quality, reduced spatial resolution, pose, and expression variations challenges. To evaluate the performance of the proposed method, we conduct thorough experiments and detailed analysis on a database of 7,500 images. The results demonstrate that the proposed system provides greater discriminability and robustness against such variations, as well as higher identification rates under various situations reflecting real-world scenarios, compared to other recently published work. Copyright © The Author(s) 2024.