在本論文中，我們提出照相手機的車牌偵測與辨識系統，用以解決車牌歪斜及車牌陰影造成的文字影像分割及辨識問題，提供一個可靠與準確的車牌偵測與辨識技術。我們所提出的方法有四個階段：分別為車牌定位、車牌校正，文字分割及文字辨識。 針對車牌定位，我們分析影像邊界點的垂直與水平方向累積量，利用邊界點累積量的數量與位置特徵分割車牌影像。車牌校正方面，我們使用區域二值化法，克服車牌陰影問題，並分析二值化影像中的白色與黑色像素點數量藉以作為判斷影像反相依據，接著利用車牌文字連結區塊bounding box的長寬比、大小以及其位置等特徵刪除非車牌區域；由於車牌傾斜的影像存在旋轉、歪斜問題，故我們將利用仿射轉換中的旋轉、歪斜轉換法，校正傾斜的車牌。 文字分割方面，我們利用邊界點垂直方向累積量與波谷分割文字，並將文字大小正規化至40×90。最後我們利用正規化相關係數樣板比對法作為文字辨識的方法，為了縮短比對時間，我們改變車牌文字辨識的程序，首先將樣板與待測影像尺寸縮小為1/4進行第一次的樣板比對，在比對結果中選取3個分數最高，進行最終的原尺寸 (40×90) 樣板比對程序，分數最高者即為所選。 This paper presents an approach for license plate recognition using a camera-equipped smartphone. The proposed method provides a reliable and accurate technique to solve the problem of license plate recognition caused by the skew and shadow on the license plates. There are four stages in the proposed approach: license plate location, license plate rectification, character segmentation and character recognition. In the first stage, we locate the license plate by accumulating edge points, and then analyze the edge points and accumulation associated with vertical and horizontal dimensions of the image. As to the second stage, license plate rectification, we adopt local threshold to cope with the problem of shadow on the plates first. Next step involved analyzing black and white pixels in order to decide whether to invert the image or not. The researcher tries to engage the characteristics like length-width ratio, size, and position of the bounding box in the text region to eliminate the non-text portions. To solve the rotation, skew, and scale problems of the slanted license plates in the image, we use an affine transformation to estimate the skew angle. Edge points vertical direction accumulating and trough are used to segment characters section in the third stage. We normalize the characters size to 40 × 90. Finally, criterion of normalized cross-correlation is used in the last stage for character recognition. In behalf of shortening the process time for identification, the procedure of character reorganization is improved. We shrink the samples to one-fourth the size to conduct the first identification process. Then, three highest-coefficient samples are chosen to match the original input pattern. From these three samples, the highest-coefficient one is selected as the final result.
