利用尺度空間二值化與累積梯度投影的方法應用於車牌字體的擷取與辨識

Abstract

本論文提出一個車牌字體辨識系統，此系統包含三個主要方法。第一個方法稱為尺度空間二值化，可以用來從灰階圖像上擷取字體。此方法結合了穩健的高斯差函數和動態二值化處理，從未知影像中直接擷取出車牌字體。為了使擷取的處理速度加快，本論文也提出優化的方法用以縮短計算時間。第二個方法稱為邊界投票方法，適合用來矯正字體在影像拍攝過程中所導致的幾何型變。此方法一開始假設了許多直線，然後以投票方式找出一條通過最多邊界點的直線當成邊界線。找到的邊界線可以幫助矯正字體的幾何型變，因而藉此改善辨識率。第三個方法稱為累積梯度投影方法，利用累積梯度並且轉換它們成特徵向量來識別獨立字體。這些特徵向量稱為累積梯度投影向量，被實驗證實對雜訊及照度改變是具有穩健性的。

A system consisting of three methods to deal with license plate characters recognition is proposed in this dissertation. The first method, scale-space binarization, is suitable for extracting characters from gray-level images. The method combines the robust Difference-of-Gaussian function and dynamic thresholding technique to extract the license plate characters directly. In order to speed up the extraction process, optimization methods are also disclosed to reduce the computation time. The second method, voting boundary method, is suitable for correcting characters from geometric deformation induced during capture process. It assumes many straight lines candidates and detects the best one passing through most of the edge pixels by voting. The boundary lines can be used for correcting the deformation and improve recognition rate thereby. The third one, accumulated gradient projection method, recognizes isolated characters by accumulating the gradient projection of the characters and converts them into feature vector for comparison. The feature vector is called accumulated gradient projection vector and is proven robust regardless of noise and illumination change in experiments.