一個可增強車牌字元辨識效能的多重辨識器設計

A Multiple-Classifier Design for Enhancing the Recognition of Characters in License Plates

張勝仁(Sheng-Jen Chang)

指導教授 ： 繆紹綱

中原大學/工學院/電子工程研究所/碩士(2003年)

https://doi.org/10.6840/cycu200300393

摘要

摘 要

本論文提出一套應用在車牌辨識上的多重系統，以達到高效能的表現，其中車牌自動辨識系統可區分為三個子系統：車牌定位、字串切割與補強切割及字元辨識。車牌定位是利用車牌影像中灰階度劇烈變化的特徵，來找出可能為車牌的區域，再利用車牌的幾何特性以達到初步車牌定位的目的。成功的將車牌定位後再微調四方邊界以去除發牌地和螺絲孔區，如此才是我們想要的車牌區域。字串的分割是依據字串外圍的輪廓來分割；此外，針對定位錯誤導致字串切割失敗的情況，我們提出以影像增強方法重新再找一次車牌，並以垂直投影法找出各字元的左右界以切割出影像中車牌的字元。另外，針對切割錯誤的車牌，根據後端字元辨識的結果來判斷是否屬於切割錯誤的情況，若果真如此，則再對已找出的字串作投影法切割出字元。在字元辨識方面，有鑑於單一辨識器其辨識率大多有一定的瓶頸，所以我們使用了兩種類神經網路辨識方法：類神經網路樣板比對法(Neural Network Template Matching, NNTM)以及倒傳遞神經網路辨識(Back-Propagation Neural Network, BPN)。根據多重辨識的觀念，我們有效結合兩種辨識方法而提升字元辨識率。
實驗結果顯示，在車牌定位上，我們整體的準確率為99.667%。在車牌辨識率方面，單獨用類神經網路樣板比對法的辨識率為94.841 %，而單獨用倒傳遞神經網路法則有96.006%，多重辨識器的辨識率則為98.668%，由此可見多重系統之效果。最後在AMD 1.5G的電腦下，每張車牌影像所花費辨識的時間只需要0.3~0.5秒。

關鍵字

多重辨識器 ； 字元辨識 ； 車牌定位 ； 類神經網路 ； 字串切割

並列摘要

Abstract
This thesis presents a multi system to achieve high the performance for license plate recognition. The automatic license plate recognition system can be divided into three sub-systems: license plate location finding, character string segmentation and enhancement, and character recognition. The solution to the location-finding problem is based on the characteristics of high gradient variation of gray-level pixels and the output of the solution is location candidates of a license plate. Then we utilize the geometric property of license plate to determine the true location from the candidates. After locating the license plate successfully, we properly adjust the boundary of the license plate to remove the province/city name and screw areas, resulting in a truly useful license plate area for further processing. The contour of the license plate provides key information for the segmentation of character string in the license plate. Furthermore, for the wrong segmentation due to bad license plate locations, we present a method that will find the location of the license plate again after the enhancement on the input image. We use a vertical projection method to find the left and right boundaries of each character from the character strings in the license plate image. The wrong character segmentation is identified by the outcome given in the character recognition phase. If that is the case, then we use the projection method to perform the character segmentation again. Normally, the character recognition rate of using a single classifier has its bottleneck. Therefore, we adopt two neural network recognition methods: NNTM (Neural Network Template Matching) and BPN (Back-Propagation Neuron). Based on the concept of multi-classifiers, we combine the two recognition methods efficiently and increase the character recognition rate.
The experimental results show that the accuracy on locating the license plate is 99.667%. On the overall recognition rate of the license plate, NNTM is 94.841, BPN is 96.006%, and the composite classifier integrating NNTM and BPN is 98.668%. This result proves the effectiveness of the multi system. Finally, the computation time of the entire license plate recognition system, running on an AMD 1.5G computer, is 0.3~0.5 seconds on the average.

並列關鍵字

Neural Network ； License Plate Location Finding ； Multiple Classifiers ； Character Recognition ； Character String Segmentation

參考文獻

[4]T. Natito, T. Tsukada, K.Yamada, K. Kozuka, and S. Yamamoto, “Robust Recognition Methods for Inclined License Plates under Various Illumination Conditions Outdoors,” IEEE Trans. Veh. Technol., Vol. 49, No. 6, Nov. 2000.

[9]S. Liang, M. Shridhar and M. Ahmadi, “Efficient Algorithm for Segmentation and Recognition of Printed Characters in Document Processing,”IEEE Pacific Rim Conf. on Commun., Computers and Signal Processing, Vol. 1, pp. 240-241, May 1993.

[11]R. Zunino, and S. Rovetta, “Vector Quantization for License-Plate Location and Image Coding,” IEEE Trans. on Indus. Elect. Vol. 47,pp. 159-167, No. 1, Feb. 2000.

[13]J. Park, V. Govindaraju, S. Srihari, “OCR in a Hierarchical Feature Space”, IEEE Trans. on Pattern Analysis and Machine Intelligence, Vol. 22, No. 4, pp. 400-407, Apr. 2000.

[20]J. Franke and E. Mandler,“A Comparison of two Approaches for Combining the Votes of Cooperating Classifiers,” Proc. of 11th Int. Conf. on Pattern Recognition Methodology and Systems, Vol. 2, pp. 611-614, 1992.