本論文提出一智慧型車牌自動辨識系統，利用類神經網路的強大學習能力，能快速有正確地辨識出車牌號碼。本系統主要由定位模組和辨識模組所構成。前者利用簡單的水平掃描線一次導數，找出車牌位置所在，後者使用一個先進的加權值共享類神經網路來學習與辨識車牌號碼。本系統設計的主要特點之一在於使用原本灰階車牌資訊進行辨識，以避免影像二值化後資訊的遺失以及切割的不正確。本系統使用Pentium-266 CPU以及64M RAM個人電腦做實驗，初步結果顯示：在取樣的173張車牌影像中，有172張定位正確，時間約為0.01秒。取其中119張車牌為訓練資料，其餘53張為測試用。類神經網路經訓練後，對原先訓練車牌辨識有116張正確，辨識率達97.4%；訓練資料亦有49張正確，辨識率達92.4﹪，辨識時間約為0.08秒。 In this paper an automatic intelligent vehicle license plate recognition system is proposed, based on the powerful learning capability of artificial neural networks. The system is primarily composed of two modules: the license plate locator and the license plate recognizer. The former utilizes simple first derivatives for horizontal scan lines to locate license plate areas. The latter, then, applies a shared-weight neural network to learn and recognize vehicle identification numbers. One of the major strengths of the proposed system is that it uses original gray-scale images of license plates for recognition in order to avoid the problems due to information loss after thresholding and incorrect character segmentation. The proposed system was tested with 173 diverse car images by a personal computer of Pentium-266 CPU with 64 M RAM. The preliminary results show that the system successfully locates 172 license plates of the experimental car images in 0.01 second/image. Among these car images, 119 are used as the training data, and the remaining 53 as test data. After training the neural network, 116 license plates of the training images are correctly recognized, archiving 97.4% recognition rate; 49 of the test images are correctly recognized, archiving 92.4% recognition rate. The recognition phase takes 0.08 second/image on an average.
公元二000年台灣智慧型運輸系統國際研討暨展覽會論文集(上冊)=Proceedings of the Taiwan's International Conference and Exhibition on ITS 2000，頁419-431