Robust Face Recognition in Low-Light Environments Using Thermal Imaging and Deep Learning Techniques

Traditional face detection systems are all primarily based on visible light. They have a very poor rate in low illumination or no illumination at all. To counter the problemsas described above, we have come up with a high-powered face detection system that comes from the principle of thermal imaging technology and captures infrared radiations emitted from the human body, thus ensuring that there is a guaran- teed successful detection even in cases of scarce illumination. Our methodology commences with detailed data preprocessing, which includes cleaning the data to remove all invalid images and proper labeling of matched and mismatched thermal and visible image pairs to enable ease in binary classification. The normalized pixel values into the range [0, 1] have enhanced the model's stability during training. Furthermore, segmentation techniques have been applied along the facial regions; these improved identification accuracy and reduced the input size goinginto the CNN. The Siamese network architecture centred on our system, composed of two identical subnetworks that learn similarities through contrastive loss, achieves validation accuracy of 96.84%. The techniques applied to stabilize training against overfitting are batch normalization and dropout regularization. In general, the results of our Siamese model for determining whether two image pairs are matched or mismatched reached a remarkable accuracy of 97.69%. This eventually shows the strong potential of thermal imaging for improving face detection in poor illumination conditions and significantly contributing to the further development of cross-domain recognition systems.