Advanced Face Authentication Using Deep Learning Models

• Published in: 2023 IEEE Pune Section International Conference (PuneCon) pp. 1 - 6
• Main Authors: E L, Arjun, Kumar, Sharan, S D, Meghana, Maharana, Suresh, S, Shriram, S, Tejonidhi, Darapaneni, Narayana, Paduri, Anwesh Reddy
• Format: Conference Proceeding
• Language: English
• Published: IEEE 14.12.2023
• Subjects: Authentication; Computational modeling; Data models; Deep Learning; Face recognition; FaceNet; OpenCV; Residual neural networks; ResNet50; Robustness; Task analysis; Transfer learning
• ISSN: 2831-5022
• DOI: 10.1109/PuneCon58714.2023.10450013

This paper conducts a comprehensive study and evaluation of various face recognition models, aiming to provide valuable insights for face recognition tasks. The models under investigation include VGGFace(ResNet50), Dlib with KNN, Siamese Network, FaceNet, and Dlib with OpenCV. Each model undergoes a meticulous assessment in terms of accuracy, robustness, and computational efficiency. The research commences with an exploration of VGGFace(ResNet50), a deep-learning model fine-tuned on a custom data set to evaluate its performance under diverse conditions. Next, the study delves into Dlib with KNN, lever-aging Dlib's capabilities for face detection, landmark detection, and encoding, combined with the K-nearest neighbors (KNN) algorithm for classification. The Siamese Network architecture is also investigated to address one-shot learning challenges, followed by an evaluation of FaceN et's highly discriminating face embedding. Finally, the integrated approach combining Dlib and OpenCV libraries for real-time face recognition is thoroughly analyzed and implemented. The findings offer crucial insights into the strengths, limitations, and applicability of each model, facilitating informed decision-making for specific face recognition applications. By combining multiple of these models, a robust FR model is built having an accuracy of more than 90 % which will play a significant role in the computer vision domain.