The Research Of License Plate Recognition System Based On Maximally Stable Extremal Regions

License plate recognition system is one of important parts of modern intelligent traffic system. After analyzing the existing research work, a novel algorithm system, which is based on Maximally Stable Extremal Regions(MSER), has been proposed in this thesis. The main work are listed as follows:First, the motivation, algorithm and implementation of MSERs is analyzed. Then based on the characteristic of license plate characters, a serial of rules have been designed to remove duplicate and noise MSERs. Thus the quantity of MSERs can be decreased and the efficiency of the algorithms can be improved.In the module of license plate location, a set of rules, based on both geometric relation between MSERs and the characteristic of license plates, have been put forward to find the candidates of license plates. Then a linear SVM classifier is applied to judge whether each MSER in the candidates of license plates is an character, and the results will be used to identify the license plate areas. Finally, an approach has proposed to rank the obtained license plate areas, and this will be helpful in the module of character segmentation. Different from the previous methods, the proposed algorithm combines both geometry and content knowledge hidden in license plates, hence its performance is improved much.In the module of character segmentation, according to the rank of license plate areas, different strategies have been designed to segment characters, which takes account of speed and accuracy of the whole system. For the first-class plates, the characters are extracted by directly using MSER information. For the second-class plates, the binaryzation, calibration of plate incline and projection segmentation are processed sequentially to extract characters. Different from the existing approaches, MSER information is utilized in each process. First, a MSER-based binaryzation algorithm is proposed, which can overcome the problem that the characters cannot be extracted under non-uniform illumination. In the process of calibration of plate incline, the incline angle is computed by fitting a straight line crossing the centers of MSERs which have been indentified as characters, and then it is used to calibrate the plates. Finally, a projection segmentation method utilizing MSER information is proposed to extract characters. The experiment results demonstrates that the proposed algorithms can accurately and efficiently extract characters of license plates.In the module of character recognition, through experiment comparison, HOG features are selected represent characters. For digit and alphabet characters, a hierarchical strategy has been proposed to recognize characters, which considers both accuracy and speed. For Chinese characters, because there is only one Chinese character in each license plate, a nonlinear SVM classifier with strong recognition ability is applied.Finally, one thousand plate images under different weather and different lighting conditions are tested. License plate recognition accuracy is 95.1%, the cost of average time is 90 ms. The result shows that the system is able to meet the requirements of real-time and accuracy.