ADVANCED TECHNIQUES FOR OCCLUSION DETECTION AND CLASSIFICATION IN FACIAL RECOGNITION SYSTEMS

Performance degradation under partial or complete occlusion of key facial regions, caused by items such as face masks, scarves and sunglasses, remains one of the most persistent obstacles facing modern facial recognition systems. Overcoming this limitation matters for the dependable operation of security, surveillance, access control and authentication platforms that must function in unconstrained, real-world settings. This research introduces a hybrid facial recognition framework built around two complementary mechanisms: Canny edge-based preprocessing, which sharpens structural facial contours, and a targeted data augmentation pipeline, which increases the diversity of training examples through synthetic and geometric transformations. Together, these mechanisms were designed to strengthen the model's reliance on occlusion-resistant structural cues rather than raw pixel intensity. The framework was tested on two widely used public benchmarks, Labeled Faces in the Wild and the AR Face Database, augmented with artificially generated masks, sunglasses and scarf occlusions to mirror real-world conditions. A convolutional neural network trained from scratch served as the recognition backbone, taking single-channel grayscale or edge-map images as input. > Recognition performance was evaluated using accuracy, precision, recall, F1 score, False Acceptance Rate (FAR), and False Rejection Rate (FRR). > Across all occlusion types, the proposed pipeline outperformed a baseline model trained directly on raw images, with average recognition accuracy rising by more than 10 percentage points alongside parallel gains in precision, recall and F1 score. Average FAR decreased from 6.8% in the baseline model to 3.4% in the proposed framework, while FRR decreased from 9.1% to 5.7%, demonstrating improved verification reliability. Because the same network architecture was used in every condition, these gains can be attributed specifically to the preprocessing and training strategy rather than to added model complexity. The results indicate that pairing edge-based structural enhancement with augmentation offers a lightweight yet effective route to occlusion-robust facial recognition, with clear relevance for deployment in real-world security and authentication contexts.