Research On License Plate Location And Character Segmentation In Complex Scenes

As the key technology of Intelligent Transportation System(ITS), the license plate recognition(LPR) is widely used. Generally, LPR is composed of the following three modules: license plate location(LPL), character segmentation(CS) and character recognition. In this thesis, LPL and CS will be mainly introduced.In actual application scenarios of the LPR, The existing methods can't simultaneously solve backlight, over-exposed, complex background, disturbance of strong noise, etc. beause of degradation of features of edge and color. As for the CS, traditional methods extract character regions under the condition of confirming four bounds of license plate and skew correction, and ignore the possibility that the result of the LPL is incomplete.Aimed at the shortage of tranditional LPL in complex application scenarios, this thesis presents a method of license plate location with good robustness. Firstly, the method blocks image, extracts histogram and computes edge density of every image block, and discriminates scene type of the image with logical regression model; According to the type, improves the image quality with special image processing methods; generates plate candidates by connection component analysis and rule filtering based on rectangle features of license plates; color segmentation is used to assist to locate; At last, excludes false regions with plate classification. Experiment tests images collected from various complex scenes, and the method finnally gets the average detection rate of 95.56%. The result shows the proposed method is effective and robust to locate license plates in complex scenes.Moreover, a new CS method based on Stroke Width Transform(SWT) is proposed. In order to ensure the integrity of the license plate, extend the result of LPL, then detect some character candidate regions with SWT, cluster to reserve real character regions based on horizontal position and height of candidate regions, and fit straight lines. According to the result of fitting lines, locate precisely position of plates, and correct the tilt plates. At last, segment characters using template matching. Experiments show that the method can accurately segment license plate characters, has good adaptability to tilt, slightly soiled plates, and solves to the problem of multi-scale plates.