- 學位論文
具備多重辨識器並以PDA為平台之 車牌辨識系統
A License Plate Recognition System Using Multiple Classifiers and a PDA Platform
摘要
車牌辨識系統的相關研究，不論動態或靜態影像都已相當多，但其研究幾乎都是以個人電腦為平台，並在固定點(例如停車場)架設固定攝影機的方式進行。隨著智慧型行動裝置的快速進步，以及網路的蓬勃發展，搭配行動裝置平台與無線網路做即時辨識傳輸，將是未來即時影像辨識的一個趨勢。因此本論文提出利用PDA (Personal Digital Assistant)內建像機在任一遠處拍照並擷取影像自動定位車牌，再利用無線網路將車牌影像傳回後端進行進一步辨識，最後將辨識結果送到網路伺服器資料庫查詢的車牌辨識系統。 在本系統前端中，我們利用Haar一階小波轉換後所得之HL頻帶的特性並搭配車牌的幾何特性去定位車牌。而在後端字元切割部分，我們採用連通元件標示法(Connected Component)先將字元找出，並搭配投影法去切割連字的部份，如遇字元模糊不清時，則採取平滑濾波器及銳化濾波器以保持車牌字元的完整。而後端字元辨識的部份，有鑑於單一辨識器辨識皆有一定的缺點，所以我們採用倒傳遞神經網路(Back-Propagation Neural Network, BPN)為主，自組織映射圖網路(Self-Organized Map, SOM)為輔的多重辨識器辨識，根據多重辨識的觀念，有效的結合兩種辨識方法提升字元的辨識率。而在連結前端與後端系統部分，我們設計出一個簡易的網頁資料庫系統，其主要功能為儲存前端辨識完的車牌影像供後端辨識器辨識，還擁有簡易的資料庫，可以立即線上查詢車籍資料。 由實驗結果顯示，在車牌定位上，整體辨識率的準確率為92.22%。在車牌辨識率方面，單獨使用倒傳遞神經網路法的辨識率為79.03%，使用倒傳遞神經網路加上自組織映射圖網路的多重辨識器辨識率則為88.43%。整個系統可說是結合了行動裝置的可攜性與無線網路的便利性，將PDA與PC的運算量透過網頁伺服器聯繫，做妥善的分配以穫得較佳效能，為車牌辨識提供一種新型態的應用。 關鍵字：PDA、車牌定位、網頁伺服器、字串切割、字元辨識、類神經網路、多重辨識器
並列摘要
There are many research works of vehicle license plate recognition based on either dynamic or static images, but almost all of them consider the use of personal computers (PC) as their study platforms, where cameras are fixed and deployed at a fixed location such as a parking lot. With the rapid development of intelligent mobile devices and the Internet, using mobile devices and wireless networks for real-time image recognition will be a trend in the future. This thesis presents a license plate recognition system with a camera inside PDA (Personal Digital Assistant) which can take a picture in any place and locate the license plate in the picture automatically. And the license plate image will be sent back to the server wirelessly for further recognition. Finally, the recognition result in terms of license plate numbers can be compared to the numbers from a license plate database in the server. In the front-end part of the proposed system, we use the characteristics of HL subband obtained by 1-level Haar wavelet transform and the geometry feature of license plates to locate a license plate. In the character segmentation part of the back-end, the connected component method is used to extract characters from the license plate and a vertical projection method is used to separate connected characters. If characters look blurring, we use low pass and sharpening filters to maintain the integrity of characters. In the character recognition part, since using a single classifier for character recognition has its weakness, we take the Back-Propagation Neural (BPN) network as our major method and supplement it by using the Self-Organized Map (SOM) neural network. Based on the concept of multiple classifiers, we combine two recognition methods effectively to increase the successful character recognition rate. In order to link the front-end and the back-end of the system, we design a simple web server database system. The database not only can save the license plate images that need to be recognized in the back-end but also can be searched immediately for the vehicle status (e.g. stolen or not). The experiment result shows that the accuracy on locating a license plates is 92.22%. The successful recognition rate of the license plate by using BPN alone is 79.03%, and by using the composite classifier integrating BPN and SOM is 88.43%. Our system takes advantages of the mobility of PDA and the convenience of wireless network. The PDA and the PC-based server can share computing burden and obtain better performance through a web server, resulting in a new type of application for license plate recognition. Keywords: PDA, License Plate Locating, Web Server, Character Segmentation, Character Recognition, Neural Network, Multiple Classifiers