| dc.description.abstract | In this thesis, we develop a license plate recognition system on Android smart phone. The proposed method consists of five stages: image pre-processing, license plate locating, orientation correction, character segmentation, and character recognition. The major properties are: (i) Segmentation of license plates is invariant to size of license plates. (ii) Segmentation of license plates is invariant to color of cars. (iii) The license plate extraction method is based on length/width proportion and color of license plates. (iv) Segmentation of license plates tolerates to the orientation variation of license plates. (v) Similar characters are recognizable.
After bi-level thresholding, we use horizontal scan twice to segment license plates. First, the width of a license plate is assumed to be the same as the width of whole image in the first scan. Second, base on the aspect ratio of the license plate, the height found by the first scan can be used to get the ideal width for the second scan. Hence segmentation of license plates is invariant to size of license plates. This segmentation is based on license-plate characters other than plates; thus, the segmentation is not influenced by body color of cars. In addition to using the aspect ratio of license plates to filter, we also use license-plate background and character color to extract license plates.
There are three steps in resolving orientation distortion of license plates. First, using average slope of upper and lower bounds of the license plate text to correct the orientation from pan rotation. Second, using license plate types to amend the total width of character blocks. Then, to find out the rotation angle that makes the license-plate height is largest, the number of connected blocks of characters is largest. At last, we use this angle to correct the tilt rotation of license plate images.
Character segmentation is implemented by vertical scan. If the aspect ratio is wrong, the corresponding segmented block will be deleted, and then each real character block of the license plate can be retrieved.
Character recognition is done by template matching. If the detected character is digit 0, letter O, letter D, and letter Q, or digit 8 and letter B, or digit 1 and letter I, we further use special character features to recognize again. The second recognition process can reduce the wrong recognition rate of these similar characters.
| en_US |