Multilingual and skew license plate detection based on extremal regions

License Plate Detection (LPD) is an important component of many applications involving security and traffic surveillance. Despite current progress, lots of hurdles remain in the way of a robust LPD system. This is particularly true for the detection of license plates with different layouts and skew angles. In this paper, a novel LPD system is proposed for detecting English and Chinese license plates with large skew angles. The proposed system consists of three main stages: 1) a character proposal module to find candidate characters based on Extremal Regions (ERs); 2) feature extraction and classification relies on Convolutional Neural Network (CNN); 3) license plate detection by region linking. The new method much improves on the robustness of existing approaches by leaving out character segmentation. The performance of the proposed license plate localization algorithm is verified using different datasets of vehicle images, including a large field-captured dataset, a skew and tilt dataset, a 12 countries dataset and two benchmark datasets. For Chinese civilian vehicles, the accuracy of plate localization is over 98.3%.