基于先验信息和射影几何变换的车牌图像矫正*

为了便于字符分割、识别，车牌识别系统需要将车牌的倾斜图像矫正为无倾斜和旋转的标准图像。目前多数文献采用Hough变换检测倾斜角度并直接进行矫正，但存在的缺陷。为此从射影几何的观点提出了新算法：利用车牌固有的先验信息，将车牌与车牌图像之间的透视变换矫正为仿射变换，再将仿射变换矫正为相似变换，将此相似变换看做是对原车牌进行旋转和全局缩放的结果，消除其中的旋转因素，就得到一个与车牌仅差一个缩放因子的标准车牌图像，以便于后续的字符处理。

License plate image rectification based on prior knowledge and projective geometric transform

For characters segmenting and recognizing conveniently, the tilt license plate image should be converted to standard image without tilt and rotation. At present, most of papers detect tilt angle by Hough transform and rectify it directly. This paper analyzed the shortcoming of this strategy and then proposed a new correction algorithm based on the viewpoint of projective geometry. The developed algorithm was as follows: By using the given prior knowledge of license plate, converted the perspective collineation between license plate and its image to affine transformation. Then converted the obtained affine transformation to similarity transformation, which contained a rotation and global scaling. After removing the influence of rotation factor, obtained a standard license plate image, which only had a scale factor compared with real license plate and was ready to following character processing.