Automatic-License-Plate-Recognition-ALPR

Introduction

Our objective is the development of an Automatic License Plate Recognition (ALPR) system designed to identify various vehicles and their corresponding license plates. The system is intended to process input in the form of videos or images containing vehicles. The desired output entails the detection of vehicles and license plates, with the results prominently displayed on the respective video.

You will gain an overarching perspective through the following pipeline:

Pipeline

Table of contents:

1. Approach

2. Overview results

1. Approach

In this task, I incorporate the use of the YOLOv8 model, Warped Planar Object Detection Network, WPOD-NET for short(you can find their paper here and the implementation in Torch here) along with Perspective Transformation, and PaddleOCR model.

• Firstly, we commence by detecting all vehicles in the frame/image

• Next step is to identify the license plate of each vehicle. WPOD-NET exhibits the ability to recognize license plates from diverse countries and diverse types of vehicles with high confidence scores.

  Example
  Confidence score	0.9897	0.9994	0.9729

• Eventually, we rectify the license plate from the frame/image using perspective transformation and input it into the PaddleOCR to get the final license plate number. In this phase, we add the correct rules just for vietnamese license plate cases to increase the accuracy of the final license plate number, so you might want to change the correct rules for your cases or disable them.

  		52LD-6042
  Input Image	Perspective transformation	PaddleOCR

  2. Overview results

  example_output.mp4

  Conclusion

  While our system effectively detects vehicles and accurately recognizes license plates, it is noteworthy that the PaddleOCR component encounters challenges in extracting text from images, primarily influenced by variations in camera setup and environmental conditions. Consequently, it is advisable to consider the incorporation of a Deep Neural Network (DNN) model to enhance image resolution before feeding it into the PaddleOCR model. This approach aims to improve the accuracy of extracting license plate numbers.

  Furthermore, in instances where the license plate image is visually clear and legible, the PaddleOCR may still encounter difficulties in extracting the license plate number. This phenomenon is illustrated in the following example: