Data Augmentation and Data Preprocessing for Machine Learning-Based Face Recognition System

Real-world face recognition accuracy declines when facial captures occur under variable, unconstrained conditions. Factors like poor ambient lighting, off-axis poses, and partial occlusions introduce noise that interferes with deep neural network feature extraction. This paper evaluates an integrated biometric pipeline designed to stabilize recognition using systematic preprocessing and active dataset augmentation. The preprocessing block employs cascaded convolutional face detection, eye-landmark affine alignment, and photometric contrast normalization via adaptive histogram equalization and self-quotient filtering. This mapping strips environmental noise from facial images prior to embedding generation. When new identities are enrolled, inline geometric and lighting augmentations synthetically expand class representation to prevent network overfitting. The system also incorporates an active, unsupervised update loop that dynamically registers unknown users using an incremental classification algorithm. This unified pipeline provides a stable, adaptive recognition architecture that preserves high classification accuracy in unconstrained operational spaces.