Automated Face Verification System Using MTCNN And Facenet Deep Neural Networks

In this work, a fully functional automated face verification system is described, built around two complementary deep learning models: Multi-Task Cascaded Convolutional Networks (MTCNN), used for detecting faces and estimating landmark positions, and FaceNet, which produces compact 128-dimensional identity embeddings. The design specifically targets difficulties that come up in real deployments — varying lighting, non-frontal poses, and partially occluded faces. On the backend, a Flask REST service computes cosine similarity between embeddings, while the React-based frontend gives users immediate visual feedback during verification. Testing on the Labeled Faces in the Wild (LFW) dataset yielded an AUC of 0.9876, an Equal Error Rate of 0.82%, and a Top-1 identification accuracy of 99.2%. Ablation experiments showed that the MTCNN face-alignment step alone accounts for a 4.5% improvement in accuracy, cutting EER from 2.11% down to 0.82%. The entire stack runs inside Docker containers, keeping inference latency around 82 ms per image on GPU and making the setup easy to reproduce across different machines.