A Two-Stage Plate Detection and Specialised Optical Character Recognition Pipeline for License Plate Recognition, Validated on Nigerian Plates from Aerial Drone Imagery

Automatic number plate recognition systems are typically trained and validated on the specific national plate format they are deployed against, which leaves an open question of how well a model trained on an internationally diverse but country-agnostic dataset generalises to a plate format, and a viewing geometry, it has never seen. This study presents a two-stage license plate detection and recognition pipeline, a YOLOv8n object detector for plate localisation followed by a specialised optical character recognition model for text extraction, trained entirely on an international dataset containing no Nigerian plates, and evaluated specifically on Nigerian plates captured from an aerial drone platform. The plate detector achieved a mean average precision at 50 percent intersection-over-union (mAP50) of 0.969, a precision of 0.985, and a recall of 0.949 on a 2,048-image validation set. Three optical character recognition approaches were compared directly: a general-purpose OCR library, a custom-trained character detector, and a specialised plate-OCR model trained across more than sixty countries' plate formats, with the specialised model producing the most consistent results across all three. When applied to three aerial drone photographs of Nigerian vehicles taken on a university campus, containing five visible plates in total, the system correctly detected all five plates and read four of them with complete accuracy; the fifth was read with one character omitted, traced to edge-of-crop clipping rather than an optical character recognition failure. Detection confidence on the drone imagery, 33 to 65 percent, was markedly lower than on the ground-level validation set, consistent with the reduced apparent plate width that a near-vertical aerial viewing angle produces, a finding this study evaluates directly against comparable prior work on aerial and drone-based plate recognition. The result demonstrates that internationally diverse training data can support genuine zero-shot generalisation to an unseen national plate format and unseen camera geometry simultaneously, without any Nigeria-specific fine-tuning, while also identifying the specific, correctable failure mode, bounding-box edge clipping, responsible for the one partial read observed.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-16
DOI
https://doi.org/10.5281/zenodo.24273040
Primary Topic
Vehicle License Plate Recognition
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

A Two-Stage Plate Detection and Specialised Optical Character Recognition Pipeline for License Plate Recognition, Validated on Nigerian Plates from Aerial Drone Imagery

Izabo Abigo
1 links
Zenodo (CERN European Organization for Nuclear Research)
Vehicle License Plate Recognition
article

A Two-Stage Plate Detection and Specialised Optical Character Recognition Pipeline for License Plate Recognition, Validated on Nigerian Plates from Aerial Drone Imagery

Izabo Abigo
article en

Abstract

Automatic number plate recognition systems are typically trained and validated on the specific national plate format they are deployed against, which leaves an open question of how well a model trained on an internationally diverse but country-agnostic dataset generalises to a plate format, and a viewing geometry, it has never seen. This study presents a two-stage license plate detection and recognition pipeline, a YOLOv8n object detector for plate localisation followed by a specialised optical character recognition model for text extraction, trained entirely on an international dataset containing no Nigerian plates, and evaluated specifically on Nigerian plates captured from an aerial drone platform. The plate detector achieved a mean average precision at 50 percent intersection-over-union (mAP50) of 0.969, a precision of 0.985, and a recall of 0.949 on a 2,048-image validation set. Three optical character recognition approaches were compared directly: a general-purpose OCR library, a custom-trained character detector, and a specialised plate-OCR model trained across more than sixty countries' plate formats, with the specialised model producing the most consistent results across all three. When applied to three aerial drone photographs of Nigerian vehicles taken on a university campus, containing five visible plates in total, the system correctly detected all five plates and read four of them with complete accuracy; the fifth was read with one character omitted, traced to edge-of-crop clipping rather than an optical character recognition failure. Detection confidence on the drone imagery, 33 to 65 percent, was markedly lower than on the ground-level validation set, consistent with the reduced apparent plate width that a near-vertical aerial viewing angle produces, a finding this study evaluates directly against comparable prior work on aerial and drone-based plate recognition. The result demonstrates that internationally diverse training data can support genuine zero-shot generalisation to an unseen national plate format and unseen camera geometry simultaneously, without any Nigeria-specific fine-tuning, while also identifying the specific, correctable failure mode, bounding-box edge clipping, responsible for the one partial read observed.

Zenodo (CERN European Organization for Nuclear Research)
Industry, innovation and infrastructure
Vehicle License Plate Recognition
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon