Research On Multi-license Plate Character Recognition Algorithm Based On Deep Learning

With the rapid development of social economy and the rapid increase of the number of cars,intelligent traffic management has become the mainstream of today's society,among which license plate recognition is one of the core links of intelligent traffic.At present,a variety of license plate recognition system products have appeared in the market and achieved good results.But for a single license plate,when the illumination is insufficient,the background is complex,the license plate tilts and becomes deformed,the recognition effect is unsatisfactory.How to overcome these problems and make the license plate recognition system more stable and universal has become a new research hotspot.In this paper,the deep learning method which has made great achievements in the field of image recognition in recent years is studied,and the convolutional neural network is adopted to realize multi-license plate character recognition,which mainly includes three steps: license plate location,character segmentation and character recognition.Firstly,in the license plate location,due to the uneven quality of the captured images,an image enhancement method based on improved two-dimensional discrete wavelet transform is proposed to improve the image quality and obtain the license plate edge information.Then,using the license plate characters to arrange the regular texture information on the license plate background to detect the number of jumps on the edge of the license plate,to roughly locate the row of the license plate;and then,by color space conversion of a small range of license plate information based on the threshold range of the license plate background color to find the possible multi-license plate position;Finally,through a priori knowledge,the license plate is screened,and all license plates are precisely located.Practice has proved that this positioning method has high accuracy and fast speed.However,the color screening process is sensitive to light,which makes the night positioning less effective.Therefore,a machine learning based method is added to make up for this deficiency.The Adaboost algorithm is improved to train the features and realize the location of multiple license plates under night conditions without relying on the detection of color features.Secondly,character segmentation is an intermediate link.Since the shooting angle may affect the license plate tilt,a method based on corner affine transformation is proposed to correct the inclined license plate.The main idea is to find the corner points of the rectangular frame of the license plate according to the shape characteristics of the license plate,and to convert the three points of the non-collinear line into a right-angled triangle to realize the tilt correction of the entire license plate.Next,a method for segmenting characters based on vertical projection and connected domain judgment is proposed for the corrected license plate,and a single character for recognition is obtained.Lastly,it is character recognition.Designing an improved AlexNet model for character recognition by analyzing current popular convolutional neural networks such as AlexNet,VGGNet,GoogLeNet,and ResNet.Due to the large number of modules,convolution kernel size,batch size and Dropout rate of the entire network,the network performance needs to be optimized by design experiments.Then,through the open source OpenCV platform,the character recognition result is printed on the license plate image to be recognized.