LicensePlate_Yolov8_Filters_PaddleOCR

Recognition of license plate numbers, in any format, by automatic detection with Yolov8, pipeline of filters and paddleocr as OCR

The main improvement with respect to the project presented before ( https://github.com/ablanco1950/LicensePlate_Yolov8_MaxFilters) has been the use of paddleocr instead of pytesseract as well as the reduction of the number of filters, some of which although they were right in certain circumstances produced noise in other.

On 04/27/2023, a new version is incorporated that considers license plates with several lines and several license plates in one image.

Requirements:

paddleocr must be installed ( https://pypi.org/project/paddleocr/)

pip install paddleocr

yolo must be installed, if not, follow the instructions indicated in: https://learnopencv.com/ultralytics-yolov8/#How-to-Use-YOLOv8?

pip install ultralytics

also must be installed the usual modules in computer vision: numpy, cv2, os, re, imutils, parabolic

All needed modules can be installed with a plain pip

Functioning:

Download the project to a folder on disk.

Download to that folder the roboflow files that will be used for training by yolov8:

https://public.roboflow.com/object-detection/license-plates-us-eu/3

In that folder you should find the train and valid folders necessary to build the model

To ensure the version, the used roboflow file roboflow.zip is attached

Unzip the files with the test images: Test.zip and test6Training.zip, taking into account when unzipping you can create a folder inside another folder,there should be only one folder, otherwise programs will not find the test images

Model Train:

the train and valid folders of the roboflow folder, resulting from the unziping of robflow.zip, must be placed in the same directory where the execution program LicensePlateYolov8Train.py is located, according to the requirements indicated in license_data.yaml

run the program

LicensePlateYolov8Train.py

which only has a few lines, but the line numbered 7 should indicate the full path where the license_data.yaml file is located.

Running from a simple laptop, the 100 epochs of the program will take a long time, but you can always lower the cover of the laptop and continue the next day (besides, there are only 245 images for training). As obtaining best.pt is problematic, the one used in the project tests is attached.

As a result, inside the project folder, the directory runs\detect\trainN\weights( where in trainN, N indicates the last train directory created, in which the best.pt file is located), best.pt is the base of the model and is referenced in line 17 of the GetNumberInternationalLicensePlate_Yolov8_Filters_PaddleOCR.py program (modify the route, the name of trainN, so that it points to the last train and best.pt created

As obtaining best.pt is problematic, the one used in the project tests is attached,it must be adjusted the route in instruction 17 in GetNumberInternationalLicensePlate_Yolov8_Filters_PaddleOCR.py

Run the program.

GetNumberInternationalLicensePlate_Yolov8_Filters_PaddleOCR_V1.py

The car license plates and successes or failures through the different filters appear on the screen.

The LicenseResults.txt file lists the car license plates with their corresponding recognized ones.

The combination of the efficiency of paddleocr and the plate detection precision of a simple yolov8 on Roboflow images, together with an adequate selection of filters, allows us to obtain a comparison of the plates recognized with the best license plate detector I know: https: //www.doubango.org/webapps/alpr/ which is included in the attached Excel file: ComparisonWithDOUBANGO.xls, testing with the images contained in the attached Test directory

By changing the path in instruction 12,activating the directory "test6Training\images" an deactivating the path "Test" in instruction 15: In a test with 117 images, 107 hits are achieved In the same manner any other image directory can be tested (In this case, the LicenseResults.txt file must be consulted to indicate the licenses plates, if the files are not named with the license number, as in test6Training occurs, it cannot be determined if the assignment was successful automatically)

The video version is also included:

VIDEOGetNumberInternationalLicensePlate_Yolov8_Filters_PaddleOCR_Demonstration.py

operating on the attached video:

Traffic IP Camera video.mp4

downloaded from project: https://github.com/anmspro/Traffic-Signal-Violation-Detection-System/tree/master/Resources

In its execution, on the monitor screen, the detected license plates are detailed with a summary at the end.

Three files are obtained:

VIDEOLicenseResults,txt with the registration of license plates detected with a lot of noise.

VIDEOLicenseSummary.txt with the following results, which seem pretty tight as can be seen visually from the video.

A8254S,145,198.2291808128357

AR606L,10,31.03719687461853

AE670S,10,15.752639770507812

A3K96,8,25.679476976394653

A3K961,3,14.658559083938599

A968B6,5,7.775115013122559

AV6190,10,17.38904595375061

The first field is the license plate detected and the second is the number of snapshots of that license plate.

As a maximum number of snapshots of 3 has been set (LimitSnapshot=3 parameter in the program), to avoid noise, the license plate of the APHI88 car that was going faster and that only recorded one snapshot does not appear (it can be checked in the VIDEOLicenseResults.txt logging file)