(IJCSIS) International Journal of Computer Science and Information Security,Vol. 8, No. 8, November 2010
applying a series of operations. The extracted LP region isfurther divided into individual characters. For recognition, theisolated characters are then transformed and fed into a pre-trained neural network.The sequence of recognized characters is then checked with adatabase. If the car is recognized, the system would thencommand the authorization for access or for billing or for sending the number plate to a central computer for further processing depending upon the application of interest.The candidate LP region extraction is the key step in anLPR system, which influences the accuracy of the systemsignificantly. The goal of this phase is, given an input image, to produce a number of regions that have high probability of containing a license plate. In the adopted approach, theextraction of the candidate region from a set of probableregions is carried out in four steps. These are explained in thefollowing subsections.Fig. 4: Sample images from the database
License Plate Extraction
The LP extraction is the main step in the overall system.The accuracy of this step affects the accuracy of the wholesystem significantly. This phase extracts the region of interest,i.e. the license plate from the acquired image. This stepconsists of three stages:
First of all, the acquired image should be converted from acolor RGB image into a gray level image (see Fig. 5). Differentgradient-based edge detectors were implemented. In our experiments, the Canny edge detector was used. The CannyEdge Detector was shown to achieve low error rate, localizededge points, and a single point edge response (see Fig. 6). Wenoticed from the collected data, that most car images exhibitedmore horizontal lines than vertical lines.Fig. 5: The original acquired image & its gray-scale imageFig. 6: Detected
edges using Canny’s method and edge
enhancement using a thickening mask
Although the Canny’s Edge Detector proved to be the
best technique for detecting the LP region, an additionalmask was required to enhance the detected edges. Thethickening mask is used for this purpose. The thickeningoperation is the dual of morphological thinning and helps instrengthening the edge structure of the image (see Fig. 6).
Detecting the Closed Boundaries of the LP Candidate Regions
By tracing the pixels in the edge detected image, theclosed structures are considered as boundaries of candidateLP regions. Tracing is carried using a 8- neighborhoodscheme. After tracing the connected pixels, a set of pre-defined conditions is then used on these boundaries for identifying the closed contours that are LP mostly likelycandidates. The conditions we use for this purpose are:
Length and Shape Filtering: This step is based on a pixelcounting approach. In our specific application, the lengthof the LP usually ranges between 1000 and 2400 pixels(from training). Our approach is to trace all boundarieswithin the given image, and select the boundariescontaining between 1000 and 2400 pixels.
Width/Length Ratio: The standard ratio of width to heightin normal Saudi license plates is 2:1. To take intoconsiderations small angle deviations, we used a ratiorange of 1.3 to 2.7. As for long license plates, the ratio isaround five (practically, between 3 and 5.5). If the areaunder study satisfies the above conditions, it is then seenas a good LP candidate.
Black to White Ratio: Normal Saudi license plates arewhite with black characters. It is found that the black to