Need of Number Plate Recognition System
This system is advanced in surveillance of cars in parking and at toll Plazas. this will automatically generate the number of a Vehicle and that can be used on the bills or may be to monitor the usage of the parking lot by a car. This system can be extremely useful for gathering statistics on road or at a check point for custom checking or to recognize a stolen vehicle. This system takes a vehicle image of any size breaks it into smaller image pieces. These pieces are then analyzed to locate the exact location of number plate in the image. Once the area of the number plate (its x and y coordinates) is found the plate is parsed to extract the character from it. These characters are then given to the OCR module. OCR program recognizes those characters and converts them in text format.
Components of the system
Vehicle number plate recognition system has three main components in it.
1. Breaking the image into smaller pieces of images which are the high frequency parts of the original image.
2. Choosing the number plate from the image pieces returned by the above module, and parsing the plate to extract out the character part.
3. Recognizing the characters in the image pieces.
TECHNIQUE USED
Signature technique is used for the implementation of this project. Taking row wise or column wise signature of an image gives the information about the less detail and more detail areas of the image. So it becomes easy to find out the areas with high frequencies.
How Signature is Used
Signature technique helps in locating high frequency areas. If the image is binarised then most of the detail is lost from the image,leaving our area of interest more prominent.
Finding Probable Number Plate In The Image
Once the image is binarised its row wise histogram (sum of white or black pixels in each row) or signature is taken to find out which number of rows is showing ridges. These ridges are basically high frequency areas and one of these ridges will definitely be a number plate. A threshold value is used to indicate the starting and ending point of the ridge. The best results were shown by taking the average of the minimum point and the median of the row signature as the threshold. Once the ridges in the row signature of the image are obtained column wise signature of those row ridges is calculated. This will further refine the candidate image by removing those columns from the row ridge which do not possess much detail. This is done by choosing the ridges our of the column histogram of the ridge area in the row histogram.
Recognizing Number Plate From The Candidate Images
After this task, the x and y coordinates of all the high frequency pieces which are the candidates of number plate are known. As we can see that on the number plate there would be 4 to seven characters. So each character will show a ridge in the row signature of the image piece, secondly most of the information is lost because of binarising the image so only number plate area will show maximum number of ridges. Now if we take the row wise histogram of those binarised pieces we can see that number plate image shows more number of ridges as compared to any other candidate image. So image with maximum number of ridges in its row signature is chosen as the number plate.Then the same signature technique is applied to extract the numbers from the number plate image. The difference was in the threshold value. Because here we needed to pick each ridge in the histogram therefore the minimum value of the histogram was chosen as the thresholding value. And the reason is that all characters might not show ridges with equal peak (highest point in the ridge). Or a character like ‘X’ might be broken into two ridges. As it is obvious that the center of the character X will show very small peak.
OPTICAL CHARACTER RECOGNIZER - OCR
What is OCR?
The goal of Optical Character Recognition (OCR) is to classify optical patterns (often contained in a digital image) corresponding to alphanumeric or other characters. The process of OCR involves several steps including segmentation, feature extraction, and classification.
Character Recognition: OCR By 2D Correlation
Description of method: Given the chosen license plate and the coordinates that indicate where the characters are, we begin the OCR process by 2D correlation. Correlate each character with either the alphabet or the numeral templates then choose the value of each character based on the result of the correlation. The first three characters on the standard License Plates are alphabets; therefore, correlate each one of them with the 26 alphabet templates. The latter four characters on the standard License Plates are numerals; therefore, we correlate each one of them with the 10 numeral templates. The result OCR is chosen based on the maximum values of the correlation for each character.
However, realized that certain characters are frequently confused. As a quick solution, we implanted a scheme to display possible alternatives. Those characters that are identified as easily misinterpreted are subjected to a correlation comparison with an array of other characters that is historically known to be easily confused with the originals. If there is a possibility that a letter or number can be confused with more than one character, the characters are listed in the output in the order of decreasing likelihood. This likelihood is based on the correlation values of the character with the various templates, where high correlation denotes a good possibility. However, this scheme only occurs if the given letter does not have a very high correlation value (does not land above a nominal threshold).
Rationale
OCR by 2D correlation is the option that seems to strike the best balance between performance and difficulty in implementation. Details can be observed in the character isolation portion of the source code.
Possible problems/Weaknesses
OCR by 2D correlation is sensitive to the size of the license plate, which meant bigger or smaller alphabets and numbers in the picture. The 2-d correlation was very sensitive to this and frequently gave back wrong results due to different size license plates.
Template matching
The steps for this process are:
1. Build a template for each of the letters to be recognized. A good first approximation for a template is to the intersection of all instances of that letter in the number plate. However, more fine-tuning of this template must be done for good performance.
2. Erode the original image using this template as structuring element. All 1 pixels in the resulting image correspond to all matches found for the given template.
3. Find the objects in the original image corresponding to these 1 pixels (e.g., using4 the function 'bwselect' in MATLAB). Another way of doing this is to implement Step 3 as a closing operation.
Advantages.
• Translation-invariant
• Simple and easy to implement
Disadvantages
• Requires considerable tweaking to find the right templates
• Not scale or rotation invariant
• Performance deteriorates rapidly for incomplete or noisy data.
No comments:
Post a Comment