A summary of “Localization of License Plate NumberUsing Dynamic Image Processing Techniques And Genetic

Algorithms"

Eric LaskaDepartment of Computer Science

University of Minnesota Morris

1. INTRODUCTION

In their paper “Localization of License Plate Number Us-ing Dynamic Image Processing Techniques And Genetic Al-gorithms”, the authors G. Abo Samra and F. Khalefah pro-vide a new evolutionary computational algorithm (geneticalgorithm) to take raw (color or grayscale) images taken of license plates and outputs the locations of license plate sym-bols. This is accomplished using a genetic algorithm theauthors developed. From here a simple optical characterrecognition algorithm has an easy task of converting theselicense plate symbols to letters, symbols, and numbers. Pre-sumably this algorithm would be used for law enforcementpurposes.

2. BACKGROUND

A genetic algorithm is an evolutionary computational al-gorithm which uses cellular genetics as a model. Given thatevolution has been successful enough to create intelligentspecies (us) using genetics, as well as practically all life aswe know it, genetics is a natural model to base evolutionarycomputation on.

Chromosomes are used to encode information into a ge-netic algorithm in a manner akin to using the letters ACGT(seen the movie GATACA?) which we use to represent thegenetic code of DNA. When we realize that this genetic codeis just like a base-4 number, we can easily extend genetic en-coding to binary, base-2, numbers.

Evolution then takes place using mutation and crossoveroperators which alter genetic letters or switch regions of ge-netic code between two individuals, respectively. We encodea goal of what we wish to evolve as logic which we may useto evaluate the fitness of genetic code, called a fitness func-tion. Once evolution has generated genes which meet orexceed some predefined fitness requirement, evolution maybe halted and a good solution is returned (although evolu-tion may still continue, fitness will level off eventually andis expected to increase slower and slower).

A population of individuals is created initially, from which

This work is licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 United States License. To view a copyof this license, visit http://creativecommons.org/licenses/by-nc-sa/3.0/us/ orsend a letter to Creative Commons, 171 Second Street, Suite 300, San Fran-cisco, California, 94105, USA.

higher-fitness members and selected, mutated, and crossed-over to produces subsequent generations. These offspringare then expected to do better than their parents when fedthrough the fitness evaluation, and the selection/crossover/mutationprocess is repeated for many iterations.

3. OUTLINE

To process a color image of a license plate into a list of li-cense plate symbols requires many steps. The authors beginby converting the image to grayscale, then converting it ablack/white-only (binary) image through a process known asadaptive binarization which highlights  the license plate sym-bols and hard boundaries. Next, a morphological operationis applied to deal with noise in the picture. A connectedcomponent analysis determines which objects in the binaryimage are the license plate characters. And finally size fil-tering is used to conclude which components are indeed thelicense plate characters when we feed all this data into agenetic algorithm. This procedure is seen in Figure 1.

Figure 1: Flowchart of localizing license plate sym-bols from an image

 

4. IMAGE PROCESSING OF PLATE

We begin with a color image of a license plate.

Figure 2: Raw color image of license plate

Next it is converted to grayscale, followed by an applica-tion of adaptive binarization, which uses neighborhood graylevels to classify pixels as foreground or background. Thishandles local variations in brightness very well.

Figure 3: Grayscale image of license plate

Figure 4: Binary image of license plate

Next, a morphological operation is applied to remove noisefrom the binary license plate image.

Figure 5: Morphological closing using 3-pixel disk-elements to reduce noise

A connected component analysis then extracts objectsfrom the processed binary image.

Figure 6: Connected component analysis of pro-cessed binary image (N=2287)

Size filtering is applied because we have a good idea of how large license plate symbols are.

 

Figure 7: Binary image after applying size filtering(M=64)

Finally a genetic algorithm developed by the authors pro-cesses the processed binary image to detect the license onthe license plate. This algorithm solves the 2D compoundobject detection problem.

• Upper left corner coordinates, width, and height of bounding boxes are encoded into chromosomes.

• Relative positions and sizes between bounding boxes,dealing with variations in skew, perspective, orienta-tion, and scaling of the image.

• Stochastic universal sampling is used in selection of off-spring. Individuals are placed along a line correspond-ing to their fitness evaluation and an equally-spaceddistribution are selected along the line. 90% of indi-viduals are exposed to evolutionary operators.

• The crossover operator used was created by the au-thors, with weights based on co-alignment of chromo-somes. This rapidly increases convergence time of evo-lution (finding a solution) by logically keeping relevantpieces of the license plate image grouped together.

• Objective distance between license plate characters isconverged to and is used in determining the correctbounding boxes of license plate characters.

The end result is a license plate with license plate symbollocations determined.

Figure 8: Binary image after applying the geneticalgorithm

5. RESULTS

This algorithm works quite well, with a 98.4% accuracy.On a 2.6Ghz PC with 2GB of RAM a 640x480 image takes0.12 seconds to locate the license plate characters, whereason a 2048x1536 image it takes 0.34 seconds. Their algorithmhas a sublinear algorithmic complexity, which is awesome.

The authors then show around a hundred instances of ap-plying their algorithm. Even with highly-variable lightingand positional conditions, as well as highly-variable licenseplate designs, their algorithm works very consistently, andin a reasonable amount of computational time.