車牌辨識系統上車牌尋找、前處理及辨識之研究

Abstract

車牌辨識系統可應用在停車場的管理系統、高速公路的收費站及贓車查緝等各種用途上。完整的車牌辨識系統由多個步驟整合而成，包括車牌尋找、影像前處理、水平軸導正處理、垂直軸字元切割及辨識五大步驟。而本論文則針對車牌尋找、影像前處理及字元辨識第三部份做詳細的分析與設計。 在車牌尋找的部份，我們利用影像粗糙化技術來降低影像的資料量，提升處理速度，再由濾波器強化車牌影像後找出車牌的位置。經由實驗結果已達到98%的正確率，相當成功。在影像前處理部份，我們依字寬條件設計濾波器來強化字型部份，再做二值化除去背景雜訊。在字元的辨識部份，我們以梯度方向的區塊統計為特徵，採用倒傳遞類神經網路對數字、英文字母和英數字混合進行辨識。字元辨識率也達90%以上。 本論文在車牌辨識技術上已達產業需求標準，但實際環境多變，需再收集應用上的問題進行改善。

The license plate recognition system could be applied to various uses, such as parking lots management system, tolls of freeways, tracking stolen vehicles, etc. A complete license plate recognition system is integrated by five main procedures, including plate localization, image preprocessing, horizontal-axis calibration, vertical-axis characters segmentation and recognition. This Thesis focuses on plate localization, image preprocessing and recognition which have been analyzed and designed in detail. In the part of plate localization, we use image roughened technique to reduce image data and improve processing speed, then plate is localized after plate image is enhanced by a designed filter. From the experiment, this method has showed very successful performance about 98%. In the part of image preprocessing, we follow the character width to design a filter to enhance the part of characters, and binarize it to eliminate background noise. In the part of characters recognition, we use regions for gradient direction histogram calculations as feature, and adopt Back Propagation Neural Network (BPNN) to recognize numerals, alphabets and alphanumerals. Characters recognition rate is above 90% The techniques of license plate recognition proposed in this thesis have achieved industrial standards. However, we need to collect and improve the using problems under various environments.