OPTIMIZED R-CNN FOR REAL-TIME PASSPORT VERIFICATION AND FORGERY DETECTION

Abstract

Manual and semi-automated passport inspections at airports cause delays and are prone to human error, especially under high passenger volumes and challenging imaging conditions. We propose a fully automated, real-time passport verification system achieving over 90% accuracy with sub-second inference on commodity GPUs. Our dataset includes 65+ country-specific passport formats from public repositories, lab captures, and synthetic augmentations. A region-based CNN with a ResNet-50 backbone and feature pyramid network detects passport regions, while machine vision extracts the Machine Readable Zone (MRZ) for checksum validation. Evaluation metrics include precision, recall, F1 score, mean average precision (mAP) at IoU thresholds 0.50 and 0.75, and latency. The system processes images in 0.75s on average, attaining 95% accuracy and outperforming classical OCR pipelines by 13 percentage points in mAP@0.50. False acceptance rates remain below 1% under variable lighting, occlusion, and print artifact conditions. Ablation studies show geometric and color augmentations improve accuracy from 88% to 95%, with diminishing returns beyond 1024×768 input resolution. Improvements are statistically significant (p < 10⁻⁴, Cohen’s d > 0.8). This work demonstrates a robust, efficient passport verification solution integrating multi-scale detection, MRZ validation, and optimized inference, paving the way for fully autonomous smart gate ecosystems with multilingual MRZ parsing and face-passport matching.