在本論文中，我們發展用於路口查驗贓車的車牌辨識系統。整個系統包括了車牌定位、車牌分類、字元切割、及字元辨識。主要的特色有：(i) 適應性二值化。(ii) 利用字元的相關性做車牌定位。(iii) 利用車牌的顏色做車牌分類。(iv) 字元寬度相關的字元切割。 在適應性二值化中，為了因應環境光線的變化，我們利用拍攝影像取樣點的灰階平均來選擇二值化門檻值。而車牌定位，因為車牌的字元間都有一定的間距，所以我們利用車牌字元的相關聯性來做定位。也就是用車牌字元邊的連續性來找影像中車牌位置。車牌的分類，我們先找出車牌背景與字元的像素位置。然後再分別利用車牌背景像素與車牌字元像素的RGB各別平均值、RGB各別最大值、與整個車牌影像的灰階平均值判斷車牌背景與字元的顏色，並做車牌分類。車牌字元切割，主要是利用連結區塊 (connected component) 並參考車牌類別做字元切割。當切割出來的字元寬度大於一個字元寬度時，則判斷此字元影像有兩個字元以上的連結，我們會按照其連結字元的寬度來平均切割。 In this thesis, we develop a license plate recognition system for finding out the stolen vehicles. The proposed system consists of four stages: license plate location, license plate classification, character segmentation, and character recognition. The major properties of the proposed system are (i) Adaptive bi-level thresholding. (ii) License plate locating based on the short horizontal distances between characters in the license plate. (iii) License plate classifying based on the colors of characters and background in the license plate. (iv) Character segmenting based on the width of characters on the license plate and the types of license plates. In the adaptive bi-level thresholding, we utilize the average of gray levels which were sampled from images to define the threshold value for bi-level thresholding. The characters on a license plate have a fixed distance, so we utilize the property to locate license plates. This means that we could utilize the clustering property of the horizontal scanning lines between characters in a license plate to detect the license plate. In the license plate classification, we extract the colors of pixels on characters and background. Then we calculate the average and maximum RGB values of the background pixels and the average gray levels of the character pixels and background pixels, respectively. We utilize those values to classify the license plates. In the character segmentation, we utilize the connected components and license plate types to segment characters. If the width of a segmented character is greater than the pre-defined width statistics of characters, the segmented character will be separated into right and left parts.
