Vietnamese Automatic License Plate Recognition with YOLO and OCR

Project Summary

NhapMonAI is a Group 05 course project for Vietnamese automatic license plate recognition. It combines a YOLO plate detector, OCR processing with FastALPR and fast-plate-ocr concepts, a Python desktop application, and a LAN-based PlateGate demonstration. The repository preserves the notebooks, checkpoints, source code, reports, presentation, screenshots, and versioned release assets used to review the work.

Field	Details
Course	Introduction to Artificial Intelligence
Faculty	Faculty of Electronics and Telecommunications, VNUHCM - University of Science
Class context	22DTV_CLC1, Group 05
Status	Completed academic prototype and public portfolio archive
Main stack	Python, PyTorch, Ultralytics YOLO, OpenCV, FastALPR, fast-plate-ocr, FFmpeg, Tkinter, Jupyter Notebook, Kaggle, Typst
Maintainer	Luong Hai Long, Student ID 22207056
Stable package	Latest GitHub release

Scope and Contribution

This is a seven-member academic project. Luong Hai Long co-developed the Python desktop application and PlateGate PC demonstration, maintained the Kaggle training evidence, and prepared the public repository, release package, and visual documentation. The repository does not assign unverified individual contributions to other members.

The system is an academic prototype, not a production traffic-enforcement product. Reported metrics apply to the project validation set; they do not establish OCR accuracy or operational reliability under every camera, plate, weather, or lighting condition.

System Pipeline

flowchart LR
  A[Image or video] --> B[YOLO plate detector]
  B --> C[Plate crop]
  C --> D[OCR and text normalization]
  D --> E[Annotated result]
  D --> F[PlateGate LAN demo]

Loading

Stage	Engineering Function
Detection	Locates candidate license plates and returns bounding boxes
Cropping	Extracts plate regions for OCR processing
Recognition	Reads candidate text and applies project normalization rules
Desktop output	Displays and saves annotated image or video results
Gate demonstration	Sends a plate string to a local allow-list workflow over HTTP

Quantitative Results

The continuation-training checkpoint reports the following detector results on the project validation set:

Metric	Result
Precision	0.99448
Recall	0.99373
mAP50	0.99450
mAP50-95	0.77006

The detector metrics are not OCR metrics. Review the notebook outputs, confusion matrices, and report discussion before comparing these values with another dataset or deployment.

Visual Evidence

Python Desktop Application	Video Inference Output

Continuation-Training Curves	Detector Confusion Matrix

OCR Demonstration	Checkpoint Continuation Evidence

Mobile Scan Demonstration	PlateGate PC Demonstration

Repository Guide

NhapMonAI/
|-- Academic_Deliverables/             # Assignment and seminar presentation files
|-- AppPythonPlateGatePC/              # PlateGate PC LAN demonstration
|-- AppPythonYOLO_OCR/                 # Desktop app, checkpoints, FFmpeg, notebooks, and outputs
|-- assets/                            # English, line-free GitHub visuals
|-- Group5_BaoCaoNhapMonAI/            # Typst report source, bibliography, and figures
|-- Group5_Notebook_IPYNB/             # Initial and continuation-training notebooks
|-- HinhAnhBaoCao/                     # Report and presentation screenshots
|-- Group5_BaoCaoSeminarNhapMonAI.pdf  # Seminar report
|-- Typst_Guide.pdf                    # Typst reference used with the report workflow
|-- v65.pt                             # Detector checkpoint
|-- RELEASE_NOTES.md                   # Current release scope and verification record
`-- README.md

Large checkpoints, application archives, and FFmpeg executables are managed with Git LFS. Install Git LFS before cloning.

Reproduce the Desktop Demonstration

1. Clone the complete repository

git lfs install
git clone https://github.com/lhlizdabezt/NhapMonAI.git
cd NhapMonAI
git lfs pull

2. Install the Python dependencies

cd AppPythonYOLO_OCR
python -m pip install --upgrade pip
python -m pip install -r requirements.txt

3. Start the desktop application

python Group5_AppPython_YOLO_OCR.py

Windows users may instead run AppPythonYOLO_OCR/run_Group5_AppPython_YOLO_OCR.bat.

4. Start the PlateGate PC demonstration

cd AppPythonPlateGatePC/PlateGatePC
python Group5_AppPYMoRongThucTe.py

The local demonstration exposes GET /health and POST /scan on port 8765. Its bien_so_duoc_phep.txt file provides the project allow list.

Invoke-RestMethod -Uri "http://127.0.0.1:8765/health"
Invoke-RestMethod -Uri "http://127.0.0.1:8765/scan" -Method Post -ContentType "application/json" -Body '{"plate":"59A22256","score":0.99,"source":"manual-test"}'

Training Evidence and References

Resource	Purpose
Group5_Notebook01_FirstTraining.ipynb	Initial YOLO training run
Group5_Notebook02_ContinuationTraining.ipynb	Continuation training from a checkpoint
group-5-vietnamese-license-plates-detection-lhl.ipynb	Kaggle-style training and output trace
Kaggle profile	Original training workspace; this repository keeps the stable notebook copy
Project dataset folder	Group dataset storage; access depends on sharing settings
FastALPR	ALPR framework reference
fast-plate-ocr	OCR component reference

Paths in the notebooks reflect the original Kaggle or local environment and may require adjustment on another computer.

Reports and Release Assets

Artifact	Review Purpose
Seminar report PDF	Project objective, method, results, and limitations
Seminar slide deck	Course presentation artifact
Python app archive	Desktop inference application package
PlateGate PC archive	LAN demonstration package
Motion GIF	Line-free animated project summary
Review card SVG	ASCII-safe detector and evidence summary
Portfolio SVG	English project overview for stable embeds
Source snapshot	Full tracked repository state, including Git LFS-backed project assets

Limitations

• Recognition quality depends on plate visibility, image resolution, viewing angle, lighting, occlusion, and plate layout.
• Dataset access and notebook paths can require permission or local path changes.
• The PlateGate component demonstrates a local LAN workflow; it is not a hardened access-control service.
• The repository has no open-source license. Treat it as a public academic portfolio archive unless a license is added later.

Frequently Asked Questions

Is this a production ALPR system?

No. It is an academic detector, OCR, desktop inference, and LAN demonstration project.

Where should a reviewer begin?

Open the latest release, review the metric table and evidence gallery, then inspect the continuation-training notebook and seminar report.

Why are the GitHub visuals line-free?

The SVG and GIF assets avoid decorative connector paths so labels remain unobstructed at desktop and narrow widths.

Are all large files included?

Yes. The tracked repository contains the project subfolders and academic artifacts; Git LFS stores the designated large binaries and archives.

Team

Student ID	Name	Public Record
22207043	Mai Xuan Khang	Group 05 member
22207106	Truong Quang Vu	Group 05 member
22207112	Ly Phi Hung	Group 05 member
22207063	Van Dinh Nam	Group 05 member
22207062	Tran Si Nam	Group 05 member
22207056	Luong Hai Long	Repository maintenance, Kaggle evidence, Python app contribution, PlateGate contribution, release packaging, and portfolio documentation
22207066	Le Tan Phi Pha	Group 05 member

Contact

Channel	Link
GitHub	github.com/lhlizdabezt
LinkedIn	linkedin.com/in/lhlizdabezt
Work email	luonghailong.work@gmail.com
Student email	22207056@student.hcmus.edu.vn
Phone	+84 988 114 708
Facebook	facebook.com/wageseadrake
Instagram	instagram.com/lhlizdabezt
YouTube	youtube.com/@lhlizdabezt
TikTok	tiktok.com/@wageseadrake