Research On Critical Recognition Technology For Low-qualified License Plate Characters

Vehicle license plate recognition is a specialized computer vision system, which has integrated many application areas such as image processing, pattern recognition and artificial intelligence. License plate character recognition is the critical parts of the whole system. It directly affects the entire accuracy and efficiency of the system. Considering the environmental situations, the characters images collected are low-qualified images, font style changes, stroke, rupture and it is difficult to recognize them correctly. So it is necessary to research an effectively character recognition method with robust to identify low-qualified characters of the license plate.After doing research on license plate character images with low quality and existing recognition method, this article proposed a multistage classifier, using different techniques to design different classifiers, according the features of letters, numbers, and characters. It contains:There exist many noise and pollution in the characters images collected. We cannot get a good result if directly put them into the classification. In this paper we did some pre-processing to improve the characters data and inhibition the unnecessary interference.As for letters and numbers, we proposed a multistage classifier combined with critical regions and the shape context features. In the first level of classification (we named it rough classification), confusing characters were classified as one class. The rough classification need reach a high recognition rate and have fast speed. The features were extracted using kernel principal component analysis and the initial classification results reached by support vector machine. If the result was not a character belonging to any confusing character pairs, the identification result was reached. Otherwise the character was sent to second level classifier (we named it refined classification).In the second classification we need to recognize the confusing characters with similar structure. The confusing characters usually differ from some certain regions, so we propose the critical regions idea. We used feature description method of the shape context, which is a descriptor with rich information of the character, taking the critical regional into account, and got the recognition results by feature matching.For characters in the license plates, we analyzed the characteristics of vehicle license plate characters and technologies on feature extraction and recognition. This paper we adopted HOG feature extraction and classified the characters using support vector machines.A high recognition accuracy rate of97.1%for numbers and letters and a record accuracy rate of91.74%for characters are obtained, showing that the proposed method is practically useful and significant especially for low-qualified characters.