運用動態資訊與多重辨識器於道路行進車輛之車牌辨識系統

Using Motion Information and Multiple Classifiers in a License Plate Recognition System for Moving Vehicles on the Road

李勝隆(Sheng-Long Lee)

指導教授 ： 繆紹綱

中原大學/工學院/電子工程研究所/碩士(2005年)

https://doi.org/10.6840/cycu200500433

摘要

本論文提出一套應用在道路中行進車輛之車牌的辨識系統，此系統包括動態影像邊緣偵測、行進車體影像擷取、車牌定位、車牌傾斜矯正、字元切割及字元辨識等子系統。在動態影像邊緣偵測上，先利用每張畫面的誤差量去偵測移動的物體，再由邊緣資訊量的大小判定是否有車體進入畫面中，若有則將此影像取出，進行車牌定位的處理。車牌定位是對邊緣資訊量最密集的地方，利用水平垂直投影法計算邊緣資訊，準確的找出車牌四個角落的座標。在字元切割上，本文採用連通元件標示法將字元找出，如字元模糊不清，則採取平滑濾波器及銳化濾波器以保持車牌字元的完整性。在字元辨識方面則採取以倒傳遞類神經網路(Back-Propagation Neural Network, BPN)為主，自組織映射圖網路(Self-Organized Map, SOM)為輔的多重確認方式及投票比對系統來達成更好的辨識率。
實驗結果顯示，字元切割的準確率接近95%。在車牌辨識率方面，單獨用BPN為81.33%，BPN搭配SOM的多重辨識器的辨識率則約90.78%，由此可見多重辨識系統之效果。最後在P4 2.8G的電腦下，每張含車牌影像所花費辨識的時間只需要0.3~0.5秒。

關鍵字

車牌定位 ； 多重辨識器 ； 類神經網路 ； 移動偵測 ； 字元辨識 ； 字串切割

並列摘要

This thesis proposes a license plate recognition system for moving vehicles on the road. The system includes the following subsystems: motion edge detection, image extraction of moving vehicles, license plate locating, tilt plate correction, character segmentation and recognition. Motion edges are detected by calculating the image difference between adjacent frames. If enough amount of difference is detected, moving vehicles are assumed in the frames and the license plate locating procedure is activated. The license plate could be identified by finding vertical and horizontal projections in the area with highly concentrated edge densities to find out the precise location of the license plate. A connected component approach is used to extract characters from the license plate. If characters look blurring, we use low pass and sharpening filters to maintain the integrity of characters. In the character recognition part, we take the Back-Propagation Neural (BPN) Network as our principle method, and Self-Organized Map (SOM) as the auxiliary method. They are combined through a voting mechanism. In this way, we could achieve better recognition performance.
The experimental results show that the accuracy on character segmentation is 95%. On the overall recognition rate of the license plate, using BPN alone is 81.33%, and using the composite classifier integrating BPN and SOM is 90.78%. This result proves the effectiveness of the multi recognition system. Finally, the computation time to perform the recognition tasks for an image containing license plate, running on a P4 2.8G computer, is 0.3~0.5 seconds on the average.

並列關鍵字

Neural Network ； License Plate Locating ； Character Recognition ； Multiple Classifiers ； Motion Detection ； Character Segmentation

參考文獻

[10]張勝仁，一個可增強車牌字元辨識效能的多重辨識器設計，私立中原大學電子工程系研究所論文，2003。

[4]T. Horprasert, D. Harwood and L. S. Davis, “A Statistical Approach for Real-Time Robust Background Subtraction and Shadow Detection,” http://www.vast.uccs.edu/~tboult/FRAME/Horprasert/ .

[7]陳麗奾，在未設限環境下車牌的定位與辨識，國立台灣大學資訊教育研究所碩士論文，2000。

[14] Y. S. Huang and C. Y. Suen,“A Method of Combining Multiple Classifiers - A Neural Network Approach,”Proc. of IAPR Int. Conf. on Computer Vision and Image Processing, Vol. 2, pp. 473-475, Oct. 1994.

[16]F. Lampariello and M. Sciandrone,“Efficient Training of RBF Neural Networks for Pattern Recognition,”IEEE Trans. Neural Networks, Vol. 12, No. 5, pp. 1235-1242, Sept. 2001.