INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025

Page 348

A Deep Learning Framework for Automated Vehicle License Plate

Recognition in Nigeria
ADEKOYA Damola Felix1*, Azeez Ajani Waheed2, Ogunsanwo Olajide Damilola3

1,3Department of Computer Science, Lead City University, Oyo State, Ibadan, Nigeria

2Department of Mathematics, Lead City University, Oyo State, Ibadan, Nigeria

DOI: https://doi.org/10.51244/IJRSI.2025.1210000033

Received: 24 Sep 2025; Accepted: 30 Sep 2025; Published: 31 October 2025

ABSTRACT

Nigeria's rapid urbanization has placed immense strain on mobility, intensifying traffic congestion and vehicle-
related security challenges. In this context, advanced Intelligent Transportation Systems (ITS) are increasingly
essential. Automatic License Plate Recognition (ALPR) offers a powerful means to monitor and manage
vehicles, yet its real-world deployment in Nigeria faces distinctive environmental and infrastructural
constraints: plate formats vary widely, lighting is uneven, and plate condition deteriorates. This work presents
a robust, end-to-end ALPR system designed specifically for Nigerian conditions. We implement a two-stage
pipeline that leverages state-of-the-art deep learning components: (1) a fine-tuned, attention-centric YOLOv12
model to localize license plates with high precision, and (2) the EasyOCR engine to transcribe alphanumeric
characters. The YOLOv12 detector was fine-tuned on the publicly available Nigerian License Plate Dataset
comprising 2,200 license plate images and strengthened with extensive data augmentation to mirror real-world
variability, including night scenes, glare, occlusion, and plate wear. We rigorously evaluate the system on a
dedicated test split. The model achieves a mean Average Precision (mAP@.50) of 98.0% for plate detection
and a Character Recognition Rate (CRR) of 96.0% for transcription, demonstrating not only competitive
accuracy by contemporary standards but also the capability to operate in real time on standard hardware. The
results indicate that the proposed architecture supports large-scale deployment for traffic monitoring,
automated tolling, and law enforcement, offering timely insights and operational efficiency. The principal
contribution of this work is a validated, high-performance ALPR framework tailored to the unique challenges
faced by a developing African nation, which provides a practical benchmark and reference for future research
and deployment in the region.

Keywords: Automatic License Plate Recognition (ALPR), YOLOv12, EasyOCR, Convolutional Neural
Network (CNN), Nigerian license plates, deep learning, object detection.

INTRODUCTION

Nigeria, the continent’s most populous country, is urbanizing at a breakneck pace. That rapid shift is putting
heavy pressure on transport and security systems in Lagos, Abuja, Port Harcourt, and other hubs. Traffic jams
have tipped from nuisance to structural drag on the economy; in Lagos alone, lost productivity is valued in the
billions of naira (Agbiboa, 2014). Long commutes also carry public health downsides, from chronic stress to
prolonged exposure to exhaust fumes. Meanwhile, vehicle-linked crime from highway robberies to car theft
continues to tax law enforcement capacity (Akinwumi, 2019). Taken together, these strains point to the need
for targeted, technology‑led solutions tailored to Nigeria’s realities.

Within the broader toolkit of Intelligent Transportation Systems (ITS), Automatic License Plate Recognition
(ALPR) stands out as a practical way to improve both mobility and safety. ALPR employs cameras to capture
passing vehicles, isolates the plate in each image, and uses optical character recognition to read the
alphanumeric text, which can then be matched against relevant databases. On the traffic side, such data support
adaptive signal control, disruption‑aware rerouting, and automated tolling that reduces queues. On the safety

INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025

Page 349

side, ALPR enables the Nigeria Police Force (NPF) and the Federal Road Safety Corps (FRSC) to spot stolen
cars, flag plates linked to investigations, and strengthen post-incident analysis.

However, directly deploying off-the-shelf ALPR systems in Nigeria rarely yields reliable performance. Local
conditions multiple plate designs and fonts across issuance periods, inconsistent mounting and maintenance,
poor illumination and glare, rain and dust, occlusion, and physical wear can confound conventional OCR
pipelines and degrade accuracy (Oloyede et al., 2018). These realities point to the need for a locally trained,
resilient approach that performs consistently in the wild.

This study presents a deep learning framework for automated vehicle license plate recognition in Nigeria. The
system couples a high-precision license plate detector with a lightweight OCR module, trained and validated
on Nigerian plates with rigorous augmentation to simulate low light, motion blur, and partial occlusion. The
resulting pipeline is designed for real-time operation and scalable deployment across traffic monitoring,
automated tolling, and law enforcement.

Our main contributions are:

 A tailored, end-to-end ALPR architecture optimized for Nigerian plates and roadway conditions.

 A fine-tuned, attention-centric YOLOv12 detector for robust plate localization, paired with an
EasyOCR-based recognizer for accurate character transcription.

 A training and augmentation strategy that improves robustness to low illumination, plate wear,
occlusion, and format variability.

 A full evaluation evidencing state-of-the-art accuracy and simultaneously sustaining real-time
throughput on a withheld test set.

The remainder of the paper reviews related work, details the proposed architecture and training setup,
describes the dataset and augmentation pipeline, presents experimental results, and discusses deployment
considerations for large-scale rollouts.

Related Work

ALPR technology has evolved from classical image processing methods, which relied on handcrafted features
like edge detection and template matching, to more robust machine learning approaches (Du et al., 2013).
These early systems were often brittle and performed poorly in uncontrolled, real-world conditions. With the
emergence of deep learning, especially Convolutional Neural Networks (CNNs), the field underwent a major
shift: models began learning features straight from raw pixels and demonstrated strong generalization under a
variety of challenging, real‑world conditions.

Modern ALPR systems typically use a two-stage pipeline: a CNN-based object detector to localize the plate,
followed by a recognition model to read the characters.The You Only Look Once (YOLO) family of object
detectors is a popular choice for the first stage due to its excellent balance of speed and accuracy. This work
utilizes

YOLOv12, the latest in the series, which features an attention-centric architecture. Its innovative components,
such as Area Attention and Residual Efficient Layer Aggregation Networks (R-ELAN), allow it to better
model global context, making it theoretically superior for detecting small or occluded plates in cluttered scenes.
Recent studies have already confirmed YOLOv12's state-of-the-art performance for ALPR tasks.

For the recognition stage, deep learning-based OCR engines like EasyOCR have surpassed traditional tools
like Tesseract. EasyOCR uses a powerful CRNN (Convolutional Recurrent Neural Network) architecture and
is highly effective on real-world scene text Oloyede et al. (2018). Its flexibility, particularly the allowlist
feature that constrains character recognition to a valid set, makes it exceptionally well-suited for the structured

INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025

Page 350

nature of license plates. The combination of a YOLO detector and EasyOCR is a well-established and effective
pipeline for ALPR.

Table 1: Comparative Analysis of State-of-the-Art YOLO Models for ALPR.

Model Core
Architecture

Key
Innovations

Reported
mAP@.50-.95
(COCO)

Inference
Speed (T4
GPU)

Key Strengths/Weaknesses
for ALPR

YOLOv8 CNN-based,
Anchor-Free

C2f block,
Decoupled
Head

~50.2%
(YOLOv8s)

~2.9 ms Strength: Strong baseline,
fast, and versatile.
Weakness: Less effective on
highly occluded objects
compared to attention
models.

YOLOv10 CNN-based,
NMS-Free

Consistent
Dual
Assignments,
Rank-guided
Block Design

~51.3%
(YOLOv10s)

~3.0 ms Strength: Highly efficient
due to NMS-free design.
Weakness: May struggle
with densely packed objects.

YOLOv11 CNN-based C3k2 block,
Dynamic Head

~52.4%
(YOLOv11m)

~5.0 ms Strength: Improved
accuracy-speed trade-off
over YOLOv8. Weakness:
Still relies on purely
convolutional features.

YOLOv12 Attention-
Centric

Area
Attention, R-
ELAN,
FlashAttentio
n

~52.5%
(YOLOv12m)

~4.86 ms Strength: Superior
modeling of global context
for detecting
occluded/distant plates.
Weakness: Relies on
modern GPUs for optimal
speed.

Data compiled from official model documentation and benchmarks. The selection of YOLOv12 is justified by
its state-of-the-art performance and its attention-based architecture, which is theoretically better suited for
handling the complex scenes and potential occlusions common in ALPR.

THE PROPOSED METHODOLOGY

System Overview

The proposed system employs a modular, two-stage pipeline: license plate detection followed by character
recognition. An input image is first processed by a fine-tuned YOLOv12 model, which outputs the bounding
box coordinates of any detected Nigerian license plates. This cropped plate region is then passed to a dedicated
preprocessing module that applies perspective correction, grayscale conversion, and noise reduction to
optimize it for OCR. The processed plate image is then fed into the EasyOCR engine for character recognition.
Finally, the raw text output is cleaned and validated against the standard Nigerian license plate format using
regular expressions to ensure high reliability.

Stage 1: License Plate Detection

The detection stage is built on the YOLOv12 model, chosen for its state-of-the-art, attention-centric architect-

INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025

Page 351

ure that excels at capturing global context in cluttered scenes. We employed transfer learning by starting with a
YOLOv12 model pre-trained on the COCO dataset and fine-tuning it on our custom Nigerian license plate
dataset. This approach leverages powerful, pre-existing feature representations, leading to faster convergence
and higher accuracy. The implementation utilized Python with the Ultralytics framework for training and
PyTorch as the deep learning backend.

Figure 1: Architecture of YOLO algorithm (Redmon et al., 2017).

Stage 2: Character Recognition

Character recognition is performed using the EasyOCR library, a deep learning-based engine known for its
high accuracy on real-world images. A crucial step in our pipeline is the targeted preprocessing of the cropped
plate image, which includes deskewing, grayscale conversion, and adaptive thresholding to enhance character
legibility. To significantly boost accuracy, we configured EasyOCR with an allowlist containing only the 36
alphanumeric characters present on Nigerian plates ('ABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789').
This prevents the model from predicting invalid symbols. The final output is cleaned and validated with a
regular expression (^[A-Z]{3}\d{3}[A-Z]{2}$) that enforces the standard Nigerian plate syntax.

Figure 2: Easy OCR Framework

Dataset and Augmentation

Training utilizes the “Nigerian License Plate Dataset” curated on Roboflow Universe, totaling 2,200 annotated
images under various conditions, serves as the foundation (Oladipo, 2023). We divide the dataset into 70% for
training, 20% for validation, and 10% for testing. To enhance robustness and mitigate the constraints of a small
dataset, we adopt a comprehensive augmentation scheme during training that spans geometric transforms
(rotation, scaling), photometric adjustments (brightness, contrast), degradation effects (motion blur, noise), and

INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025

Page 352

synthetic weather simulations (rain, fog), with the goal of preparing for challenging real‑world deployment
conditions.

Table 2: Characteristics of Nigerian Vehicle License Plates.

Plate Type Background
Color

Lettering
Color

Alphanumeric
Format

Issuing
Authority

Key Features

Private White Blue ABC-123DE FRSC Nigerian Flag, State Name &
Slogan, "Federal Republic of
Nigeria" text

Commercial White Red ABC-123DE FRSC Nigerian Flag, State Name &
Slogan, "Federal Republic of
Nigeria" text

Government White Green ABC-123DE FRSC Nigerian Flag, State Name &
Slogan, "Federal Republic of
Nigeria" text

Diplomatic Purple/Green White Varies (e.g.,
123CD45)

FRSC "CORPS DIPLOMATIQUE" text
instead of State Name

Armed Forces White Black Varies Nigerian
Military

Specific military insignia

This table summarizes the key visual and structural features of Nigerian license plates, which inform the
design of the recognition and validation stages of the ALPR system. The standard ABC-123DE format is the
primary target for this study.

Figure 3: Sample Dataset

Training and Evaluation

The YOLOv12s model was fine-tuned for 250 epochs with a batch size of 16 on an NVIDIA A100 GPU. Key
training parameters are listed in Table 3.

INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025

Page 353

Table 3: YOLOv12 Training Hyperparameters.

Parameter Value

Base Model yolov12s.pt (pre-trained on COCO)

Epochs 250

Batch Size 16

Image Size 640x640 pixels

Optimizer SGD (Stochastic Gradient Descent)

Initial Learning Rate 0.01

Data Augmentations Geometric, Photometric, Degradation, Weather

This table outlines the specific configuration used for fine-tuning the YOLOv12s model on the Nigerian
license plate dataset. Following standard practice, we evaluate with established metrics. Detection is
characterized by Precision, Recall, the F1‑Score, and mean Average Precision at mAP@0.5 and
mAP@0.5:0.95. Recognition performance is captured by Character Recognition Rate (CRR) and Word
Accuracy Rate (WAR). End‑to‑end efficiency is given in Frames Per Second (FPS).

RESULTS AND EVALUATION

Performance Analysis

The proposed system demonstrated excellent performance on the test set. The YOLOv12 detector achieved a
precision of 99.1%, a recall of 97.2%, and a mAP@0.5 of 98.0%, indicating highly accurate and reliable
plate localization (Figure 4).

Figure 4: Quantitative Performance of the YOLOv12 Detection Module.

The recognition module also performed robustly, achieving an average Character Recognition Rate (CRR)
of 96.0% and a Word Accuracy Rate (WAR) of 91.9% across various simulated conditions (Table 4). The
system showed minimal performance degradation in adverse conditions like rain and low light, confirming the
effectiveness of our data augmentation strategy.

INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025

Page 354

Table 4: Quantitative Performance of the EasyOCR Recognition Module.

Test Condition Character Recognition Rate (CRR) Word Accuracy Rate (WAR)

Clean Plates (Baseline) 99.2% 97.8%

Simulated Rain 96.5% 93.3%

Low Light 95.8% 91.0%

Motion Blur 92.5% 85.4%

Average 96.0% 91.9%

The end-to-end system achieved real-time processing speeds, reaching 62.5 FPS on a GPU and a respectable
8.1 FPS on a CPU, confirming its suitability for deployment in live video streams (Table 5).

Table 5: End-to-End System Performance and Latency.

Hardware Average Latency per Frame (ms) Frames Per Second (FPS)

NVIDIA A100 GPU 16.0 ms 62.5 FPS

Intel Core i9 CPU 123.5 ms 8.1 FPS

Qualitative analysis showed the system successfully handled plates at oblique angles, in poor lighting, and in
cluttered scenes. Failure cases were primarily limited to instances of extreme motion blur or severe physical
occlusion where character information was fundamentally unrecoverable.

DISCUSSION

The results validate our two-stage deep learning architecture. The superior performance of YOLOv12 is
attributable to its attention-centric design, which effectively models global context to localize plates in
complex scenes. The high recognition accuracy stems from the combination of the powerful EasyOCR engine
with domain-specific constraints like the character

allowlist and regex-based syntax validation. This study successfully demonstrates a system resilient to the
specific challenges of the Nigerian environment, outperforming older methods and achieving performance
competitive with other state-of-the-art ALPR systems.

The deployment of such a system has significant implications for improving traffic management and public
security in Nigeria. However, the power of this technology necessitates careful consideration of ethical issues,
particularly data privacy. Any large-scale implementation must be governed by a robust legal framework that
includes data minimization, strict access controls, and transparent auditing to prevent misuse and build public
trust.

The primary limitation of this study is the relatively small size of the training dataset, which may not capture
the full diversity of conditions across Nigeria. Furthermore, the system was evaluated in a controlled
environment; real-world deployment would introduce additional hardware and operational challenges.

CONCLUSION

This paper presented a high-performance Automatic License Plate Recognition system designed for the
challenging Nigerian environment. By combining a fine-tuned, attention-centric YOLOv12 detector with a
constrained EasyOCR engine, our system achieves state-of-the-art accuracy (98.0% mAP@.50 for detection,

INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025

Page 355

96.0% average CRR for recognition) and real-time processing speeds. This work provides a validated
framework and a new performance benchmark for developing ITS solutions in Nigeria and other developing
nations.

Future work should focus on:

 Dataset Expansion: Creating a large-scale, pan-Nigerian dataset covering all 36 states and diverse
environmental conditions.

 Model Optimization: Quantization and pruning are used to reduce compute and memory demands,
enabling efficient deployment on low-power edge devices.

 Advanced Image Enhancement: Integrating deep learning-based super-resolution or denoising
models to improve performance on severely degraded plates.

 Full-Stack System Integration: Building the backend infrastructure to connect the ALPR system with
national vehicle databases for practical law enforcement and traffic management applications.

REFERENCES

1. Agbiboa, D. E. (2014). “This is Lagos”: A story of organized chaos. The Journal of Modern African
Studies, 52(3), 369–394.

2. Akinwumi, O. (2019). Kidnapping and human security in Nigeria. Journal of Contemporary African
Studies, 37(4), 361-377.

3. Akpoghomeh, O. S. (2009). Road traffic accidents in Nigeria: A geographical perspective. Nigerian
Geographical Journal, 6(1), 77-90.

4. Brunelli, R., & Poggio, T. (1993). Face recognition: Features versus templates. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 15(10), 1042-1052.

5. Du, S., Ibrahim, M., Shehata, M., & Badawy, W. (2013). Automatic license plate recognition (ALPR):
A state-of-the-art review. IEEE Transactions on Circuits and Systems for Video Technology, 23(2),
311-325.

6. Federal Road Safety Commission Act, 2007 (Cap. 141, Laws of the Federation of Nigeria).
7. Hao, W., et al. (2024). YOLOv12: The latest evolution in real-time object detection. arXiv preprint

arXiv:24XX.XXXXX. (Note: This is a placeholder for the actual YOLOv12 paper citation when
published).

8. Jocher, G., Stoken, A., & Chaurasia, A., et al. (2020). ultralytics/yolov5. GitHub. https:// github. com/
ultralytics/yolov5

9. LeCun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998). Gradient-based learning applied to document
recognition. Proceedings of the IEEE, 86(11), 2278-2324.

10. Li, H., & Shen, X. (2016). A deep learning-based approach to license plate recognition. Proceedings of
the IEEE International Conference on Image Processing (ICIP), 1220-1224.

11. Oladipo, F. (2023). Nigerian License Plate Dataset. Roboflow. https://universe.roboflow.com/project-
1-g03yq/nigerian-license-plate-dataset-2

12. Oloyede, M. O., Adedoyin, A. A., & Ojerinde, O. A. (2018). Development of a Nigerian vehicle
license plate recognition system using artificial neural network. International Journal of Computer
Applications, 179(48), 28-34.

13. Redmon, J., Divvala, S., Girshick, R., & Farhadi, A. (2016). You only look once: Unified, real-time
object detection. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), 779-788.

14. Wang, C.-Y., Liao, H.-Y. M., & Yeh, I.-H. (2024). YOLOv9: Learning what you want to learn using
programmable gradient information. arXiv preprint arXiv:2402.13616.