Thermal image identification against pose and expression variations using deep learning

Thermal infrared face recognition systems have developed as an effective complement to visual systems for biometric identity purposes and military operations. With the help of deep learning and convolutional neural networks, we provide an efficient approach for thermal facial identification that can handle position fluctuations and expression dissimilarities in this study. As an essential deep learning model that has demonstrated its effectiveness in several computer vision and machine learning applications, we employ the ResNet-50 architecture which consists of 50 layers of convolution, activation, and pooling. The structures of those layers are discussed in detail to gain a profound insight about the operation of this architecture. In this regard, a deep and detailed mathematical analysis is furnished. The system is implemented on a dataset of 1500 thermal images, where we execute experiments in various setups and circumstances to address the issues with posture and expression variance. The experimental results show that the system achieves an accuracy rate of 99.4% when it is trained using 30% of the dataset after five epochs. With respect to other performance measures, the system attains 100% recall, precision, F-score, and specificity. In comparison to recently published works, the findings show that the suggested system offers improved discriminability, resilience against fluctuations, as well as high identification rates under diverse settings that mimic real-world scenarios.