EGP-YOLO: A Lightweight License-Plate Detector with Ghost Convolution for Edge Deployment
Abstract
Automatic license-plate detection is a core component of intelligent transportation systems, and is increasingly required to run on embedded cameras rather than in the cloud. Modern nano-scale detectors, however, still allocate a large share of their parameters and computation to standard convolutions, which strains the storage and power budgets of edge devices and leaves limited headroom for auxiliary modules such as low-light enhancement. We present EGP-YOLO, a lightweight single-stage detector that reconstructs the backbone and neck with Ghost convolution and inserts a position-sensitive attention block at the deepest stage. Because Ghost units generate part of each feature map through cheap linear operations, the model approximately halves the convolutional cost while the attention block restores localization accuracy under the tighter budget. On a public license-plate benchmark, EGP-YOLO uses 1.97 M parameters and 5.2 GFLOPs, reductions of 34.6% and 35.8% over the YOLOv8n baseline, while retaining a competitive mAP50:95 of 49.7%. Under half-precision deployment its size drops to 4.11 MB and GPU power to 80.4 W with no accuracy loss, and a CLAHE-based cascade extends it to low-light recognition.