Efficient real-time license plate recognition using deep learning on edge devices
Fedi Sonnara, Hamadi Chihaoui, Fethi Filali
Abstract
Abstract Real-time automatic license plate recognition (ALPR) is essential for smart-traffic, tolling, parking, and policing, yet roadside cameras must run on $$<\!10$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mrow> <mml:mo><</mml:mo> <mml:mspace/> <mml:mn>10</mml:mn> </mml:mrow> </mml:math> W hardware with limited memory and patchy connectivity, ruling out cloud off-loading. These constraints demand compact, fast models resilient to oblique views, motion blur, glare, and diverse plate styles. We introduce Light-Edge , a single-pass deep network that jointly localizes plates and recognizes characters. It shares a ResNet-18 + FPN backbone, removes 28 % of convolutions with a $$1\times 1$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mrow> <mml:mn>1</mml:mn> <mml:mo>×</mml:mo> <mml:mn>1</mml:mn> </mml:mrow> </mml:math> channel-fusion block, and replaces anchors with an anchor-free head followed by a CTC decoder. After mixed-precision compilation in Torch-TensorRT, the 38 MB model sustains 14 FPS on a Jetson Nano—73 % faster than the anchor-free AF-Net (8.1 FPS) and 49 % faster than YOLOv8-MobileLPR (9.5 FPS)—while keeping competitive accuracy (90.2 % mAP) and halving AF-Net’s power consumption (4.8 W vs 8.8 W). Light-Edge therefore satisfies the stringent speed–accuracy envelope required for large-scale, privacy-preserving ALPR on resource-constrained edge devices.