Technical Application Note 001

Guidelines for Image Acquisition for ANPR Sample Data and Applications

Revision:1.3
Date:2024-03-06
Contact:support@carrida-technologies.com
Copyright:1996-2023 CARRIDA Technologies GmbH, Ettlingen, Germany
Author:CARRIDA Support

Home page

Table of Contents

1   Introduction

This document describes the properties of sample images needed for the training of the CARRIDA OCR image. It also serves as a reference for optimal image aquisition for the regular use of CARRIDA.

The CARRIDA automatic number plate engine needs to be trained for each country or state to be recognized. For this purpose a large number of license plate image samples is required so that its machine learning algorithms can be used zo their full potential.

2   Desired image sample properties

In order to train a new CARRIDA classifier, at least 2.000 images should be provided, and of those samples, no more than 4 images of a unique license plate should be acquired.

The more images are provided, the better the quality of the resulting detector. Countries with license plates of different fonts require more samples to cover all font types.

The following basic image properties are required for a successful training of the |car| classifier:

  • All images must be sharp and without motion blur.

    Hint

    Hint Sign If using a video camera or digital photo camera, switch to manual mode, then set the shutter time to 3-5 ms, and manually adjust the f-stop and ISO setting to capture a properly exposed image.

  • No interlace in video frames, no half frames.

  • The background of license plates shall not be overexposed, a slightly underexposed image is preferred over an image which is too bright.

  • In order to achieve a certain diversity in the sample image set, at least 3 different viewpoints shall be used for image acquisition of the samples. This is not a strict requirement, but will increase the quality of the classifier.

  • A mixture of street scenes and highway scenes is preferable.

  • Since generally ANPR is done with cameras at a viewing position higher than a typical license plate, the training images should be taken with the camera placed at least 1 m above ground.

  • License plates should not be rotated more than 20 degrees.

    ./images/imageTan1.png
  • The aspect angle of the camera relative to the car rear side or front side (sideways viewing angle) shall not exceed 30 degrees left or right, up or down.


    ./images/imageTan2.png

  • The imaging resolution should allow at least a 1 pixel gap between characters in the license plate.

  • The minimum width of the vehicle should be around 400 pixels. For example, in images with VGA resolution (640x480 px) the width of a car should be more than half the width of the image.

  • The minimum height of characters in an image shall be at least 14 pixels for training samples (for special cases see below).

./images/imageTan12.png

Example for the max and min character sizes, and vehicle size.

3   License plates with special characters

Some countries enhance their license plates with special characters and markers to provide additional information about their use. For example China, Thailand, UAE, Saudi Arabia, the USA, and Russia enhance their plates with symbols, state signs, small strings or stacked characters.

The resolution of the images for license plate reading in those countries must consequently be higher than usual in order to render those special characters clearly visible.

We recommend an image resolution which results in a minimal character height of at least 25 pixels. In this case the size of the special symbols will be big enough so that they become easily readable.


./images/imageTan7.png

Good sample images (left side), and bad sample images (right side) for license plates from UAE, Thailand and China. The resolution in the images on the right side is insufficient.


4   Example Images

The following images are good examples for the specifications and recommendations given in this document.

i1 i2 i5
i3 i4 i6


Images displayed below are bad examples. Some are not sharp enough, some plates are covered by other objects or the characters are too small for recognition.

b1 b2 b5
b3 b4 b6