One-Shot Face Recognition Using Siamese Neural Networks: A Deep Learning Approach

The growing demand for secure identity verification has driven advancement in facial recognition technology. Traditional face recognition system often require large dataset for training, particularly when data is limited. The general challenge in face recognition lies in the need on large labeled datasets to train reliable models which can be a barrier in many real-world scenario like secure access, identity verification, or personalized services, collecting numerous images per individual is often unfeasible. As a result there is a need for systems that is capable of accurate identification from a single image. The study compares three Siamese Neural Network-based models for face verification tasks, focusing on face recognition with minimal data. Model A is the baseline, using a standard Siamese Neural Network using CNN for feature extraction. Model B enhances Siamese Network using ResNet18 as the backbone, improving feature extraction and performance. Model C combines these improvements with a triplet-loss-inspired forward pass using a modified MobileNetV3 as the backbone, enhancing the model’s ability to distinguish between similar and different faces. All models were trained and evaluated using a synthetic face dataset. Model C demonstrated superior performance in differentiating positive and negative pairs, while Model A showed limited accuracy. This results highlights the potential of one shot learning in face recognition system, especially in areas where security and data limitations are a concern.