應用倒傳遞網路模型設計動態即時車牌辨識系統

Using Back Propagation Model to Design a Dynamic Real-Time License Plate Recognition System

陳彥任(Yen-Ren Chen)

指導教授 ： 黃有評

大同大學/電機資訊學院/資訊工程學系所/碩士(2006年)

摘要

車牌辨識系統的應用十分廣泛，舉凡停車場自動化管理、收費站自動化、贓車查緝都包括在內。過去的研究大都是著重在靜態系統的研究，需要人為方式先拍好清楚方正的車牌，然而，這種方式對車輛的位置和移動性以及環境都已事先做了條件限制，本篇論文提出動態即時車牌辨識系統的實作方案，可以對一連串的影像畫面做處理，同一車輛可做一次以上的辨識來增加成功機率。此類動態系統的拍攝方式，盡可能不被人為所限制，較適合應用在真實環境中。
本研究首先從監控系統錄製的影片來源取出影像畫面，判斷畫面有車子經過時，進行梯度分析，利用水平垂直投影整合搜尋框的方式找出車牌位置，然後判斷傾斜角度進行轉正；接下來，去除雜訊邊界(車牌上的螺絲孔、螺絲及發牌地)，分割出車牌字元的個別區塊，針對每一個區塊判斷是否含有破折號或是包含兩個以上的字元，加以分析處理並切割字元；最後，將分割出來的個別字元區塊做正規化，分別產生水平與垂直投影的直方圖資料，利用倒傳遞網路模型進行文字辨識。依台北市監理處公佈的號牌規範說明，以破折號為分隔，只有兩碼字元的部份為英文字母與數字的組合，四碼字元的部份只包含數字。因此，為了增加辨識的正確性，將這兩個部份各自代入已訓練完成的權重值進行字元比對。本論文以實際在省道錄製的車輛影片進行即時車牌辨識系統的測試，並針對所提出的車牌辨識技術詳加說明，藉由測試的結果以驗證所提方法之有效性。

關鍵字

即時車牌辨識 ； 梯度分析 ； 水平和垂直投影 ； 搜尋框 ； 倒傳遞網路模型

並列摘要

License plate recognition systems have been used extensively in many applications, such as in the automation of parking lots and tollgates, as well as in investigations of stolen vehicles. Most research focuses on static systems, which require a person to take a clear and even image of the license plate. However, the ability to obtain a usable image with this type of system can be restricted by the location or movement of the target vehicle and by the clarity of the environment. We present a scheme to implement a real-time license plate recognition system to process a series of video frames, allowing for a car to be recognized more than once in order to increase the probability of success. By using already existing surveillance cameras, the real-time system does not have the same restrictions as a static system, allowing it to be applied to a wider variety of environments.
The system works by first obtaining the video frames from previously captured video recordings. If a frame includes passing vehicles, we can process the frame by performing gradient analysis and take advantage of horizontal and vertical projections and a search window to find out the location of the license plate within the frame, and then estimate the angle of inclination of the license plate to adjust the image. Secondly, the system removes the noise boundary (consisting of holes, screws, and location text on the license plate), dividing the image into several individual character blocks. Each block is checked to find if it contains dash or more than two characters, and is then analyzed and segmented again if necessary. The final steps are to normalize the segmented characters, create horizontal and vertical histogram data, and use this data to identify the characters using a back propagation neural network.
According to the license plate standards established by the Motor Vehicles Office of the Taipei City Government, a license plate has two sections, separated by a dash. The first section of a license plate has only two characters and can consist of a combination of the 26 English alphabet letters and the 10 numerical digits. The second section has four characters and can consist of only numerical digits. Therefore, to increase recognition accuracy, the template for the first section will be created by training with a different set of possible weights than the template for the second section.
This thesis uses practical video files to test and verify the system and to explain the proposed technology, as well as uses experimental results to prove the effectiveness of the proposed method.

並列關鍵字

Real-time license plate recognition ； gradient analysis ； horizontal and vertical projections ； search window ； back propagation neural network

參考文獻

[1] H.-L. Bai and C.-P. Liu, “A hybrid license plate extraction method based on edge statistics and morphology,” Proceedings of the 17th International Conference on Pattern Recognition, vol. 2, pp.831-834, Aug. 2004.

[2] H.-L. Bai, J.-M. Zhu and C.-P. Liu, “A fast license plate extraction method on complex background,” Proceedings of the IEEE Intelligent Transportation Systems, vol. 2, pp.985-987, Oct. 2003.

[3] D.-G. Bailey, D. Irecki, B.-K. Lim and L. Yang, “Test bed for number plate recognition applications,” Proceedings of the 1st IEEE International Workshop on Electric Design, Test and Applications, pp.501-503, Jan. 2002.

[4] J. Barroso, E.-L. Dagless, A. Rafael and J. Bulas-Cruz, “Number plate reading using computer vision,” Proceedings of the IEEE International Symposium on Industrial Electronics, vol. 3, pp.761-766, July 1997.

[5] G.-Z. Cao, J.-Q. Chen and J.-P. Jiang, “An adaptive approach to vehicle license plate localization,” The 29th Annual Conference of the IEEE Industrial Electronics Society, vol. 2, pp.1786-1791, Nov. 2003.