Edge-Optimized Real-Time Number-Plate Detection for CCTV using Multi-Scale Knowledge Distillation and Quantization

Abstract

Edge-based Automatic License Plate Recognition ALPR Systems are having a hard time serving the real-time detection accuracy in complex CCTVs. The state-of-the-art high-resolution solvers rely on server-class hardware with over 50 GFLOPs and 200 MB memory, which cost between USD 5000–8000 per node. On the other hand, lightweight edge models provide low detection precision (<0.75) and recognition accuracy (40%) rendering use limited in practice of traffic monitoring and law enforcement. In this paper we introduce a real-time on-the-edge ALPR system that exceeds the 25 FPS mark, running at commodity end-point devices with hardware computational budget of less than 10 GFLOPs and memory footprint below 100 MB, while yielding detection precision greater than 0.85 with recognition accuracy exceeding 0.90. The adopted method is composed of multi-scale knowledge distillation to transfer semantic and spatial information from a high-capacity teacher network to a thin student model through pixel-wise and attention-aware manners. Feature Pyramid Networks can cope with variable license plate sizes (50–200 pixels) in real-world CCTV images. Quantization-Aware Training, which effectively converts weights and activations to 8-bit integer precision, reduces the computation cost by 0.75. The dataset augmentation for the CCTV setting adds condition-dependent augmentations including motion blur (speeds of vehicles up to 120 km/h) as well as illumination changes (10 to 100,000 lux), compression, and viewing angles between ±45°.This approach provides a cost effective, privacy-preserving on-device ALPR solution that can potentially lower the deployment costs by 10–50 times as compared to server-based systems and scale intelligent transportation systems in resource-limited domains.