Research On License Plate Recognition Algorithm Under Complex Environment

With the development of economy and technology, the number of vehicles is increasing, which makes traffic problems more urgent. So Smart City and Intelligent Transportation Systems are attracting more and more attention and become a research focus. As the most important part of Intelligent Transportation Systems, License Plate Recognition technology becomes more and more important.License Plate Recognition includes five major steps:image preprocessing, license plate location, plate correction, characters segmentation and characters recognition. The existing license plate recognition algorithm now have some problems, for example, when there is motion blur phenomenon in the image or in complex background, the result of license plate location, correction and character segmentation is not ideal. In addition, the accuracy of Chinese character recognition is relatively low. In this paper, we will give deep research about these problems.(1) If image-forming system moves too fast relative to its subjects, the captured image often is blurred by motion. So we present the image blurring model, then we analysis and derivation it, and we get regular pattern when we transform the image to frequency domain. According to these, we propose a method based on twice Fourier transform to get point spread function. Firstly, we use Fourier transform twice to transform the fuzzy image to frequency domain so that we can get the angel of blurring. Then, we apply local adaptive binarization method to result of first Fourier transform, and get its projection after rotation, based on which we can compute length of blurring. Finally, improved Lucy-Richardson method is used to deblurring image, and this method can suppress the ringing effect well.(2) In reality, we will encounter some complex background, such as, poor lighting at night and rainy days, strong sunlight and the similar areas like license plate so that the accuracy of license plate location is not ideal. This paper put forward a license plate location method based on twice color calibration and connected domain analysis. This method not only adapts to a variety of complex environments, but also improves the accuracy of the license plate location.(3) In this paper, we use method of connected domain analysis to compute horizontal tilt angle, and use method of rotation and projection to compute vertical shearing angle. When we rotate and sheer image, we also need to rotate and sheer circumscribed rectangle of connected domains, which makes characters in their own rectangles. So we just need to do some simple transformation to get the result of character segmentation. Experiment shows that this method not only save lots of time, but can ensure the high accuracy.(4) Traditional method of license plate character recognition needs we design and extract character features, so if these features are unreasonable, the accuracy of license plate recognition will be affected. So in this paper, the convolution neural network model is applied to the license plate character recognition, and this model can extract features itself which can describe characters better. Experimental results show that this method improves the accuracy of character recognition, especially for Chinese character recognition.