The thermographic face model appliance for identification and authentication

The subject of this study pertains to the continued exploration of identity identification algorithms utilizing thermogram data of individuals. The novelty of this research lies in the development of a biometric thermographic model for facial recognition, capable of incorporating quasi-static points that remain unchanged over time and unaffected by external factors. Additionally, this work provides experimental confirmation of the proposed Method. Within the scope of this study, the investigated model primarily focuses on the capacity to generalize key facial areas and determine the variability weights of these areas for everyone, regardless of when the thermogram of the user's face was captured. Furthermore, this study train algorithms that have undergone testing in previous works by the authors, as well as within the current research, forming part of the developed software and hardware complex designed to validate the concept of identifying quasi-static areas. The primary outcomes of this research reveal that the developed hardware and software complex corroborates the applicability of the algorithm based on Siamese convolutional neural networks for resolving the problem of user identification via facial thermography. This algorithm successfully competes with other biometric identification methods, including those utilizing 2D facial images. Moreover, experimental results affirm the effectiveness of isolating quasi-static regions, with an 86.41% accuracy rate in correctly identifying features identified using the developed neural network. Concerning the task of amalgamating results obtained from two identification algorithms, the stacking method based on the logistic regression algorithm emerges as the most effective. This approach yields low error rates of the first and second kinds, amounting to 6.61 and 5.63%, respectively. Furthermore, in the comparative analysis of identification algorithms based on 2D images of users' faces, the FaceNet algorithm is deemed the most effective. This algorithm boasts high identification accuracy, even in the presence of alterations in appearance and lighting conditions. FaceNet proves especially advantageous when working with 2D facial images, significantly enhancing performance in user identification tasks based on images captured within the visible spectrum. The f-measure metric for this algorithm, trained on data from the YouTube Face DataBase, stands at 0.94. The practical significance of these research findings lies in their potential application within access control and monitoring systems to enhance the reliability of person-based authentication processes. The effective utilization of the methods presented here holds value in the realm of processing thermographic images for the purpose of identity verification, relying on the quasi-static features of the user's facial thermogram. Such applications can be especially valuable in scenarios necessitating identification, even amidst fluctuations in facial expressions, appearance changes (such as makeup application), or variations in environmental conditions.