利用散射變換做車牌辨識系統

本論文探討汽車車牌辨識研究，並針對台灣在 2012 年所發行的新式汽車車牌提出一個新的字元辨識方法。而提出的方法中包含車牌定位、車牌字元切割及車牌字元切割，車牌定位的部份是應用開放碼 OpenALPR，定位出車牌矩形的四點並計算 Homography matrix 來正規化車牌影像。字元切割方法是基於字元模組來設計結合字元投影輪廓及已知字元寛度的混合法，可以有效切割出字元影像。利用散射變換(Scattering Transform)演算法來擷取車牌字母之特徵值，並使用支持向量機(SVM, Support Vector Machine)模組做分類來辨識字元。為訓練SVM，我們發展能模擬車牌取像及車牌定位誤差的影像合成系統，利用合成的擬真字母影像做為訓練集。完成的車牌辨識系統對於車牌辨識有不錯的效果，並能抵抗字元影像一定程度上的型態變化，平均辨識率達98.82\%。

關鍵字

散射變換；支持向量機；車牌辨識

並列摘要

In this thesis, we proposed a new method for recognizing the Taiwanese car license plates issued since 2012. The proposed method consists of a license plate detection step, a license plate rectification step, a character segmentation step, a feature extraction step, and a character recognition step. The license plate detection step is accomplished by using the OpenALPR, an open source license plate recognition software package. The four corners of the detected license plate are used to compute a homography matrix to rectify the license plate image. A model-based character segmentation method is developed to obtain single-character images. Features of the character images are extracted with scattering transform. Characters are recognized using the support vector machine (SVM). In order to train the SVM, a license plate image synthesis system is implemented to provide the training images. The image synthesis system generates license plate images corrupted by several effects such as out-of-focus blur, shadow, image noise, and character deformation due to the license plate localization error. Real experiments show that the average accuracy of the proposed method is 98.82\%. It even outperforms the convolutional neural network approach in digits recognition.