使用監督式K平均分群法與支持向量機之階層式車牌辨識系統

Abstract

近年來運用車牌辨識技術於路口監視器吸引許多注意，它是實現智慧城市重要的一環，用來偵測肇事或遺失的車輛。過去車牌辨識已經成熟的運用在停車場的管理系統，達到停車免票證，停車位置記錄的功能。不同於停車場固定拍攝角度與光線的環境，運用於路口監視器的車牌辨識會遭遇因為拍攝角度、多車道偵測、車輛行駛速度與環境光影等因素而造成字元辨識困難。除了上述環境條件以外，字元辨識常見挑戰還包含：車牌模糊、髒汙、相異字體、相近字元等變因。本論文提出一個以SVM分類器為核心之車牌辨識系統，系統分為三個部分，包含車牌偵測、字元分割，與字元辨識。 車牌偵測的方式，使用Support Vector Machine (SVM)分類器。SVM分類器的目的為分類車牌及非車牌區域，而本研究使用以Histogram of oriented gradient (HOG)為訓練特徵的SVM分類器。為節省計算時間，過程使用圖形處理器(GPU)加速SVM計算。實驗結果顯示，我們的系統在三車道內擁有97.69%的車牌偵測成功率。抓取車牌影像後，將車牌上的字元分離，此步驟透過水平投影去除車牌上下方非字元排列之其他區域，再以垂直投影方法，分離車牌上字元。 最後字元辨識部分，本論文提出一個結合supervised K-means與Support Vector Machine的階層架構，先透過supervised K-means，將辨識字元分成子群，對於子群的字元再透過Support Vector Machine進一步分類與辨識，可以降低SVM的複雜度並提升SVM的辨識率。實驗結果顯示我們所提出的階層架構，可達98.89%之字元辨識準確率，相較於純粹使用SVM的車牌辨識技術，我們得到3.6%的辨識率改善。In recent years, the use of license plate recognition technology in traffic monitor has attracted a lot of attention because it can be used in a smart city to do criminal investigation and traffic detection. Licenseplate recognition technology has been widely used in parking lot management systems which has fixed shooting angle and lighting environments. The license plate recognition used in traffic monitor will encounter difficulties in character recognition due to factors such as shooting angle, vehicle speed and environment light and shadow. In addition to the above environmental conditions, the common challenges of character recognition include: license plate fuzzy, dirty, different fonts, similar characters and other changes. This paper presents a license plate identification system with SVM classifier as the core. The system is divided into three parts, including license plate detection, character segmentation and character recognition. License plate detection section, we use the Support Vector Machine (SVM) classifier. The purpose of the SVM classifier is to classify the license plate and the non-license plate area, and this study uses the SVM classifier with histogram of oriented gradient (HOG) as the training feature. In order to reduce the computing time, To save the computation time, we use graphics processor units (GPU) to accelerate SVM calculations. Experimental results show that our system in the three lanes with 97.69% license plate detection rate. After seizing license plates, character segmentation is adopted to separate characters. In this stage, we through the horizontal projection method to remove the other non-character arrangement om the license plate, and then use the vertical projection method to separete the license plate into characters. Finally in the last stage of character recognition, this paper presents a hierarchical architecture combining supervised K-means and support vector machine. The supervised K-means is used to classify characters into subgroups. The characters of subgroups can be further classified by support vector machine. The advantage of the proposed approach is to reduce the classes of characters in each subgroup to further reduce the number of SVMs and their complexity, and thus improve the accuracy of character recognition. Experimental results show that our proposed hierarchical architecture achieves an accuracy of 98.89% in character recognition. Compared with the license plate recognition technology using SVM alone, we get a 3.6% improvement in recognition rate.