Vietnamese Automatic License Plate Recognition with YOLO and OCR

Executive Summary

NhapMonAI is an academic computer vision portfolio project for Vietnamese automatic license plate recognition. The repository combines YOLO/PyTorch plate detection, OCR/FastALPR recognition, Kaggle/IPYNB training evidence, a Python desktop inference application, a LAN-based PlateGate PC demo, Typst report sources, seminar materials, visual assets and release snapshots.

The project is written for two audiences. HR reviewers can scan the repository quickly through the release, visuals, metrics and contact links. Engineering reviewers can trace the claims back to notebooks, source folders, checkpoints, screenshots, reports, release assets and bounded prototype notes.

Five-Minute Review Path

Step	What to Open	Why It Matters
1	Visuals at the top of this README	Shows the project scope without requiring a full source review
2	Release page	Provides a stable package of report files, demo assets and source snapshots
3	Quantitative Results	Grounds the YOLO claim in validation-set metrics
4	Evidence Gallery	Shows the desktop app, video inference output, confusion matrix and gate-demo evidence
5	How to Run	Gives a practical path for running the Python app and LAN demo

Project Profile

Field	Details
Repository	lhlizdabezt/NhapMonAI
Portfolio category	Computer vision, machine learning, OCR and Python desktop tooling
Academic context	Introduction to Artificial Intelligence course project, Group 05, Faculty of Electronics and Telecommunications
Maintainer profile	Luong Hai Long
Student ID	22207056
School	VNUHCM - University of Science
Primary stack	Python, PyTorch, Ultralytics YOLO, OpenCV, FastALPR, fast-plate-ocr, FFmpeg, Tkinter, Kaggle, IPYNB, Typst, Git LFS
Latest release	GitHub Releases
Tags	Version tags

Evidence Highlights

Evidence	Repository Material
Plate detector	YOLO/Ultralytics checkpoint and training outputs
OCR path	FastALPR / fast-plate-ocr references, plate-crop normalization and Python post-processing
Training trace	Kaggle notebook, local IPYNB files and checkpoint-continuation evidence
Desktop demo	Python Tkinter/OpenCV/FFmpeg app for image and video inference
Gate-control demo	PlateGate PC prototype with /health and /scan endpoints on port 8765
Documentation	Typst report sources, seminar PDF, PPTX slide deck, screenshots and release notes
Portfolio packaging	English README, self-hosted visuals, GitHub release, tags, topics and Git LFS rules

Role and Scope

Item	Honest Scope
Team context	Academic team project with seven members
Luong Hai Long's portfolio work	Maintained the public repository, normalized the README, release, tags, topics and visual evidence for review
Luong Hai Long's development work	Co-developed the Python desktop app and PlateGate PC demo within the team context
Prototype boundary	Academic dataset and LAN demo environment; not a production traffic-enforcement system
Review value	Evidence for applied computer vision, Python tooling, technical documentation, release packaging and engineering communication

System Overview

flowchart LR
  A["Input image or video"] --> B["YOLO plate detector"]
  B --> C["Plate crop"]
  C --> D["OCR and normalization"]
  D --> E["Annotated output"]
  D --> F["PlateGate PC LAN demo"]
  E --> G["Report and release evidence"]
  F --> G

Loading

Stage	Description
Input	User provides image or video data for plate detection
Detection	YOLO model detects plate regions and returns bounding boxes
Crop	Detected plate regions are extracted for downstream recognition
OCR	FastALPR / fast-plate-ocr workflow reads and normalizes candidate text
Output	The app renders annotated media and saves reviewable output files
Gate demo	LAN prototype accepts recognized plate strings for allow-list checking

Quantitative Results

The final reported detector checkpoint from the continuation-training stage has the following validation-set metrics:

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

These metrics describe the detector on the project validation set. They should not be read as a production guarantee for OCR accuracy, uncontrolled lighting, difficult camera angles, occluded plates or traffic-enforcement deployment.

Evidence Gallery

Python Desktop App	YOLO and OCR Video Output

Confusion Matrix	Gate Demo Evidence

Repository Structure

NhapMonAI/
|-- Academic_Deliverables/             # Seminar slide deck and assignment evidence
|-- AppPythonPlateGatePC/              # LAN PlateGate PC demo package
|-- AppPythonYOLO_OCR/                 # Python YOLO/OCR app, outputs, FFmpeg and checkpoints
|-- assets/                            # Self-hosted README visuals
|-- Group5_BaoCaoNhapMonAI/            # Typst report source, figures and bibliography
|-- Group5_Notebook_IPYNB/             # First-training and continuation-training notebooks
|-- HinhAnhBaoCao/                     # Report and slide screenshots
|-- Group5_BaoCaoSeminarNhapMonAI.pdf  # Final seminar report
|-- RELEASE_NOTES.md                   # Release summary for the current portfolio snapshot
|-- v65.pt                             # Main YOLO checkpoint
`-- README.md

How to Run

Install Git LFS before cloning because the repository contains model checkpoints, app bundles and larger media assets.

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

Run the Python desktop app for image or video inference:

cd AppPythonYOLO_OCR
python -m pip install --upgrade pip
python -m pip install -r requirements.txt
python Group5_AppPython_YOLO_OCR.py

Run the Windows batch launcher:

cd AppPythonYOLO_OCR
run_Group5_AppPython_YOLO_OCR.bat

Run the PlateGate PC LAN demo:

cd AppPythonPlateGatePC/PlateGatePC
python Group5_AppPYMoRongThucTe.py

Check the demo endpoints after the PlateGate PC app is running:

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"}'

Endpoint	Purpose
GET /health	Confirms that the local demo server is running
POST /scan	Sends a recognized plate string to the demo workflow
Port 8765	Default local LAN demo port
bien_so_duoc_phep.txt	Allow-list file used by the PC gate demo

Notebook, Dataset and References

Source	Role	Notes
Group 5 Kaggle notebook	Training and experiment trace	Paths may need adjustment outside the Kaggle environment
Dataset folder on Google Drive	Dataset storage used by the group	Access depends on the Drive sharing configuration
FastALPR	ALPR framework reference	Useful for comparing OCR and inference organization
fast-plate-ocr	OCR-related reference	Useful when comparing or replacing OCR components

Notebook in This Repository	Purpose
Group5_Notebook_IPYNB/Group5_Notebook01_FirstTraining.ipynb	First YOLO training run from scratch
Group5_Notebook_IPYNB/Group5_Notebook02_ContinuationTraining.ipynb	Continuation training from checkpoint
AppPythonYOLO_OCR/group-5-vietnamese-license-plates-detection-lhl.ipynb	Kaggle-style notebook with runnable traces and outputs

GitHub Metadata

Metadata Item	Current Intent
Repository description	Vietnamese ALPR portfolio project using YOLO/PyTorch and OCR/FastALPR, with Kaggle/IPYNB training evidence, desktop inference, LAN demo, reports and release-backed visuals
Topics	alpr, license-plate-recognition, vietnamese-license-plate, yolo, ocr, computer-vision, machine-learning, pytorch, opencv, kaggle, fast-alpr, object-detection, python, jupyter-notebook, git-lfs, tkinter, ffmpeg, typst, academic-project, hcmus
Release	Latest release
Tags	Tag list
Main languages	Python, Jupyter Notebook, Typst
Large assets	Managed with Git LFS where appropriate

Release Assets

Asset	Review Use
Seminar report PDF	Read the project objective, method, results and limitations
Seminar slide deck	Review the class presentation artifact
Python YOLO/OCR app archive	Run or inspect the desktop inference application
PlateGate PC archive	Inspect the LAN gate-control prototype
Line-free motion GIF	Visual reviewer aid with English labels and no animated connector lines
English SVG review card	ASCII-safe visual summary for profile embeds and release evidence
Line-free portfolio SVG	English profile visual with no moving dotted or curved lines
Source snapshot	Versioned repository snapshot tied to the release tag

Frequently Asked Questions

Question	Answer
Is this a production traffic-enforcement product?	No. It is an academic prototype and portfolio archive with validation-set detector metrics and LAN demo evidence.
What is the strongest engineering signal?	The project connects model training, OCR workflow, desktop inference, LAN demo behavior, documentation and release packaging.
Where should reviewers start?	Start with the release page, this README, the quantitative results table and the visual evidence gallery.
Why are releases and tags important here?	They preserve stable review snapshots for HR, faculty review and engineering inspection.
Does the README use English only?	Yes. The public-facing README text, table labels, captions and SVG-visible text are written in US English.
Why are the visuals line-free?	The current visual pass avoids animated connector lines, dotted paths and curved lines so text remains unobstructed in GitHub rendering.

Team

Student ID	Name	Portfolio Note
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, release packaging, Kaggle evidence, Python app contribution and portfolio documentation
22207066	Le Tan Phi Pha	Group 05 member

Maintainer 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

Transparency Notes

• The repository is an academic project archive and should be reviewed as evidence of technical workflow, not as a production deployment.
• Dataset paths and Google Drive access can require permission or path changes outside the original environment.
• OCR behavior depends on image quality, camera angle, lighting, occlusion, plate layout and post-processing rules.
• The repository does not currently include a formal open-source license, so treat it as a public academic portfolio archive until a license is added.

Writing Standard

This README uses an evidence-first Harvard/IEEE-style professional standard: direct technical nouns, bounded claims, reproducible links, release-backed assets and clear distinction between coursework, prototype behavior and production deployment.