Study On License Plate Detection And Recognition Algorithm From Natural Scenes

License plate is the unique identification of the vehicle, and its particularity and importance determines license plate recognition system become an indispensable important component of the intelligent traffic management. As an important research topic of information processing technology and pattern recognition, license plate recognition system also reflects the great application value in monitoring peccancy vehicles and car park management, etc. There are a lot of researches for license plate recognition, but it is a considerable challenge to detect and recognize license plate fast and exactly due to some interference factors such as the weather, lighting, camera angle and distortion in natural scene.This paper achieves a license plate recognition system from natural scenes, and it primarily consists of accurate license plate detection, character segmentation and character recognition. In license plate detection, a method from rough detection to accurate detection is proposed. Color fixed collocation feature and morphological operations are sequentially utilized to obtain candidate regions; Non license plate regions are deleted through Connected Component Analysis (CCA), and extract HOG feature of the rest region and its adjacent regions, then use SVM for the regions classification to accurate locate license plate. In character segmentation stage, a method combined CCA and inherent character is put forward. Firstly, segment numbers and letters by CCA based on binary image. Then segment Chinese character based on the width and height of character to achieve segmentation of whole plate. In character recognition stage, we design a method that includes Chinese character classifier, letter classifier, number and letter mixture classifier. Stroke feature and grid feature are computed for Chinese character recognition; Two-level classification scheme is designed for number and letter characters recognition, the method can effectively improve precision of similarity characters recognition.Moreover, our research team establishes a database that includes412images from natural scene. Experiment results on our database show that the proposed method has high precision, speed and strong robustness.