Abstract
The growing demand for real-time, adaptive facial recognition in resource-constrained environments like telemedicine, surveillance, and biometric authentication necessitates scalable AI solutions. Existing systems often falter under low-data conditions or limited computational resources. This paper introduces AIFS, an efficient and hybrid facial recognition framework that unifies traditional feature-based learning with modern few-shot deep learning under a shared Siamese architecture. The framework proposes two synergistic approaches: (1) a lightweight edge-oriented path using the Viola-Jones algorithm combined with Particle Swarm Optimization (PSO) for facial feature extraction within a Siamese network, optimized for low-power devices, and (2) a deep learning cloud-oriented path using a Siamese network with triplet loss, employing EfficientNetV2 and InceptionV3 as high-capacity feature encoders for enhanced generalization from limited examples. The proposed AIFS framework is validated across diverse platforms to simulate real-world deployment, with CPUs and Raspberry Pi representing resource-constrained edge devices, and GPUs representing high-capacity cloud environments. Tested on the Kaggle Face Recognition Dataset under a one-shot, low-data setting, AIFS achieves up to 99% accuracy. The results demonstrate a balance between latency, inference speed, and resource efficiency, confirming AIFS as a scalable and robust solution for real-time facial recognition in heterogeneous computing scenarios.