Automatic License Plate Recognition─an Implementation for Dual Character Formats

• Main Authors: Hao-En Lan, 藍浩恩
• Other Authors: Shao-Wei Leu
• Format: Others
• Language: zh-TW
• Published: 2013
• Online Access: http://ndltd.ncl.edu.tw/handle/22036623969579443738

碩士 === 國立臺灣海洋大學 === 電機工程學系 === 101 === The highway authorities in Taiwan started to issue vehicle license plates based on revised formats near the end of 2012. Although the new license plates will eventually replace all the old ones, the transition period could take years. The purpose of this thesis is therefore to address the problem of automatic license plate recognition for license plates based on two distinctive formats. Our implementation for license plate recognition is purely software. The software system includes three major components: license plate localization, character segmentation, and character recognition. The localization step includes edge detection with Sobel’s method, progressive edge noise filtering, and edge points filtering. The character segmentation step first performs binarization on candidate areas with adaptive thresholds then segments characters using connected-component labeling. Horizontal tilt correction is also performed when necessary. Characters are finally recognized by a single-layer artificial neural network. Closely resembling characters are further distinguished by their differences in grey levels of pixels at specific corresponding corners. The overall recognition rate is 89.8% based on a sample space of 717 pictures of vehicle license plates. If old plates and new plates are considered separately, the overall recognition rates are 92.2% for the old formats and 82.8% for the new formats. The successful rates for all major processing steps are 98% for plate localization (98.5% for old and 96.7% for new), 95.7% for character segmentation (96.4% for old and 93.7% for new), and 95.6% for character recognition (97% for old and 91.4% for new). The lower successful rates for the new formats are largely due to the difficulty to collect enough images of new license plates, because new license plates are currently far from being common. This also explains the less-than-satisfaction character recognition rate for the new license plates and indicates the direction for improvement.