基於非遞迴式週期性離散小波轉換於滑動式視窗車牌辨識研究

CNU IR > College of Sustainable Environment and Recreation > Dept. of Applied Geoinformatics > MOST Project > Item 310902800/24957

Please use this identifier to cite or link to this item: https://ir.cnu.edu.tw/handle/310902800/24957

Title:	基於非遞迴式週期性離散小波轉換於滑動式視窗車牌辨識研究 Based on Non-Recursive Discrete Periodic Wavelet Transform Using Sliding Window of License Plate Recognition
Authors:	蔡金豐
Contributors:	應用空間資訊系
Keywords:	滑動式視窗非遞迴式離散週期性小波轉換類神經網路
Date:	2011
Issue Date:	2012-01-03 13:57:36 (UTC+8)
Abstract:	有別於傳統的車牌辨識系統所使用之車牌定位、範圍選取、字元切割、字元辨識等複雜的處理流程，本計畫以一滑動式視窗對擷取影像進行掃瞄，並於掃描過程中同步執行小波轉換，透過一維非遞迴式離散週期性小波轉換取得所有階數之小波轉換係數，此係數經由三值化處理後作為類神經網路輸入之特徵值，經過倒傳遞類神經網路的學習訓練將輸出結果與資料庫之車牌內容比對以達到自動化車牌辨識。為降低雜訊干擾影響與加速辨識效率，在此僅選擇最後三階之小波轉換係數做為車牌辨識之特徵值，並將掃描線以間隔方式擷取，以減少計算時間但保有其辨識能力，資料庫內有100輛車牌資料，每輛都以六種不同的距離拍攝以模擬車輛行進中之動作。在類神經網路訓練與測試過程採用Leave-One-Out 方法，初步實驗統計結果，車牌成功辨識率可達93%以上，驗證本系統具有高辨識率與高容錯(fault-tolerance)等特性。 Despite the complex process of the identification system used by the conventional license-plates, such as license-plate location, range selection, character segmentation and character identification, this project utilizes sliding window to capture the image and scan. Moreover, wavelet transform is processed simultaneously during the scanning process. The 1-dimension non recursive discrete periodic wavelet transform (1-D NRDPWT) is utilized to convert coefficients and all levels of wavelet coefficients are obtained. The coefficients are under the process of tri-values and used as the character value of the artificial neural network input. The Back-Propagation Artificial Neural Network outputs results and database for license-plates comparison in order to achieve the automation of license-plate identification. In order to reduce noise interference and accelerate the efficiency of identification, only the last three levels of wavelet transform coefficients are selected as the identification value for the identification of license-plates. Furthermore, the scanning line is captured in an interval way in order to reduce the computing time and at the same time maintain the identification ability. The database contains information of 100 vehicles and each has been shot from six different distances while the model vehicles are on the move. In the training of the artificial neural network training and testing process utilize the Leave-One-Out method. The preliminary experiment results show the rate of 93% on the successful identification of the license-plates and this testifies the characteristics of the high identification rate and high fault tolerance rate.
Relation:	計畫編號：100-2221-E-041-018 計畫年度：100；起迄日期：2011-08-01~2012-07-31 核定金額：316,000元
Appears in Collections:	[Dept. of Applied Geoinformatics] MOST Project

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	2576	View/Open

Loading...