nomeroff-net: A Jupyter Notebook repository from InzGIBA

Nomeroff Net. Automatic numberplate recognition system. Version 2.1.0

Introduction

Nomeroff Net is an opensource python license plate recognition framework based on the application of a segmentation neural network and cusomized OCR-module powered by GRU architecture.

The project is now at the initial stage of development, write to us if you are interested in helping us in the formation of a dataset for your country.

Version 2.1 2.5x faster Nomeroff Net 1.0.x! This improvement was achieved by replacing segmentation model CenterMask2 with a Object Detection YoloV5 and Scene Text Detection CRAFT models.

Installation

Installation from Source (Linux)

Nomeroff Net requires Python >= 3.6 and opencv 3.4 or latest

Clone Project and clone related projects

git clone https://github.com/ria-com/nomeroff-net.git
cd nomeroff-ne

For Centos, Fedora and other RedHat-like OS:

# for Opencv
yum install libSM

# for pycocotools install 
yum install python3-devel 

# ensure that you have installed gcc compiler
yum install gcc

yum install git

# Before "yum install ..." download https://libjpeg-turbo.org/pmwiki/uploads/Downloads/libjpeg-turbo.repo to /etc/yum.repos.d/
yum install libjpeg-turbo-official

For Ubuntu and other Debian-like OS:

# ensure that you have installed gcc compiler
apt-get install gcc

# for opencv install
apt-get install -y libglib2.0
apt-get install -y libgl1-mesa-glx

# for pycocotools install (Check the name of the dev-package for your python3)
apt-get install python3.6-dev

# other packages
apt-get install -y git
apt-get install -y libturbojpeg

install python requirments

pip3 install "torch>=1.8"
pip3 install "PyYAML>=5.4"
pip3 install "torchvision>=0.9"
pip3 install Cython
pip3 install numpy
pip3 install -r requirements.txt

Installation from Source (Windows)

On Windows, you must have the Visual C++ 2015 build tools on your path. If you don't, make sure to install them from here:

Then, run visualcppbuildtools_full.exe and select default options:

Hello Nomeroff Net

# Specify device
import os

# Specify device
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
os.environ["TF_FORCE_GPU_ALLOW_GROWTH"] = "true"

# Import all necessary libraries.
import numpy as np
import sys
import cv2

# NomeroffNet path
NOMEROFF_NET_DIR = os.path.abspath('../')

sys.path.append(NOMEROFF_NET_DIR)

# Import license plate recognition tools.
from NomeroffNet.YoloV5Detector import Detector
detector = Detector()
detector.load()

from NomeroffNet.BBoxNpPoints import NpPointsCraft, getCvZoneRGB, convertCvZonesRGBtoBGR, reshapePoints
npPointsCraft = NpPointsCraft()
npPointsCraft.load()

from NomeroffNet.OptionsDetector import OptionsDetector
from NomeroffNet.TextDetector import TextDetector

from NomeroffNet import TextDetector
from NomeroffNet import textPostprocessing

# load models
optionsDetector = OptionsDetector()
optionsDetector.load("latest")

textDetector = TextDetector.get_static_module("eu")()
textDetector.load("latest")

# Detect numberplate
img_path = 'images/example2.jpeg'
img = cv2.imread(img_path)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)

targetBoxes = detector.detect_bbox(img)
all_points = npPointsCraft.detect(img, targetBoxes,[5,2,0])

# cut zones
zones = convertCvZonesRGBtoBGR([getCvZoneRGB(img, reshapePoints(rect, 1)) for rect in all_points])

# predict zones attributes 
regionIds, countLines = optionsDetector.predict(zones)
regionNames = optionsDetector.getRegionLabels(regionIds)

# find text with postprocessing by standart
textArr = textDetector.predict(zones)
textArr = textPostprocessing(textArr, regionNames)
print(textArr)
# ['JJF509', 'RP70012']

Hello Nomeroff Net for systems with a small GPU size.

Note: This example disables some important Nomeroff Net features. It will recognize numbers that are photographed in a horizontal position.

import os

# Specify device
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
os.environ["TF_FORCE_GPU_ALLOW_GROWTH"] = "true"

# Import all necessary libraries.
import numpy as np
import sys
import cv2

# NomeroffNet path
NOMEROFF_NET_DIR = os.path.abspath('../')

sys.path.append(NOMEROFF_NET_DIR)

# Import license plate recognition tools.
from NomeroffNet.YoloV5Detector import Detector
detector = Detector()
detector.load()

#from NomeroffNet.OptionsDetector import OptionsDetector
from NomeroffNet.TextDetector import TextDetector

from NomeroffNet import TextDetector
from NomeroffNet import textPostprocessing

# load models
#optionsDetector = OptionsDetector()
#optionsDetector.load("latest")

textDetector = TextDetector.get_static_module("eu")()
textDetector.load("latest")

# Detect numberplate
img_path = 'images/example2.jpeg'
img = cv2.imread(img_path)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)

targetBoxes = detector.detect_bbox(img)

zones = []
regionNames = []
for targetBox in targetBoxes:
    x = int(min(targetBox[0], targetBox[2]))
    w = int(abs(targetBox[2]-targetBox[0]))
    y = int(min(targetBox[1], targetBox[3]))
    h = int(abs(targetBox[3]-targetBox[1]))
    
    image_part = img[y:y + h, x:x + w]
    zones.append(image_part)
    regionNames.append('eu')
    
# find text with postprocessing by standart
textArr = textDetector.predict(zones)
textArr = textPostprocessing(textArr, regionNames)
print(textArr)
# ['RP70012', 'JJF509']

Hello Jupyter Nomeroff Net

Online Demo

In order to evaluate the quality of work of Nomeroff Net without spending time on setting up and installing, we made an online form in which you can upload your photo and get the recognition result online

AUTO.RIA Numberplate Dataset

All data on the basis of which the training was conducted is provided by RIA.com. In the following, we will call this data the AUTO.RIA Numberplate Dataset.

We will be grateful for your help in the formation and layout of the dataset with the image of the license plates of your country. For markup, we recommend using VGG Image Annotator (VIA)

Dataset Example:

AUTO.RIA Numberplate Options Dataset

The system uses several neural networks. One of them is the classifier of numbers at the post-processing stage. It uses dataset AUTO.RIA Numberplate Options Dataset.

The categorizer accurately (99%) determines the country and the type of license plate. Please note that now the classifier is configured mainly for the definition of Ukrainian numbers, for other countries it will be necessary to train the classifier with new data.

AUTO.RIA Numberplate OCR Datasets

As OCR, we use a specialized implementation of a neural network with GRU layers, for which we have created several datasets:

If we did not manage to update the link on dataset you can find the latest version here

This gives you the opportunity to get 99% accuracy on photos that are uploaded to AUTO.RIA project

Number plate recognition example

Contributing

Contributions to this repository are welcome. Examples of things you can contribute:

Training on other datasets.
Accuracy Improvements.

Credits

Dmytro Probachay <dmytro.probachay@ria.com>
Oleg Cherniy <oleg.cherniy@ria.com>

InzGIBA/nomeroff-net