Research On Blind License Plate Recognition Algorithm Based On Convolutional Neural Network

License plate recognition has a practical significance for the maintenance of traffic safety,urban security,traffic congestion preventing,and traffic automation management.At present,the license plate recognition in practical application has achieved great success,but the algorithm is mainly based on the fixed camera position,and the license plate to meet certain assumptions.This thesis aims at recognizing license plate whose position,size,and number are not limited,which is called blind license plate recognition,and the main works are as follows:1)A based on convolution neural network license plate detection algorithm is proposed.Firstly,character detection model is trained by convolution neural network and SVM classifier.The input image is scanned by sliding window line by line,then put the patches into the character detection model and calculate the score.Secondly,the non-maximal suppression method is used to obtain the maximum value of each row,and the position of the license plate is determined according to the position between the maximum values.Experimental results show that using a single character to train convolution neural network can improve the detection accuracy of the blind license plate effectively.2)A based on convolution neural network license plate recognition algorithm is proposed.Firstly,numbers and letters are trained together and Chinese characters are trained respectively,.the models are composed of convolution neural network and multi-class SVM.Then,the detected license plate is divided into a Chinese character region and a non Chinese character region.In the Chinese character region,the Chinese characters model is used to recognize the Chinese characters,and in the non Chinese characters region,the number and letter recognition model is used to recognize the numbers and letters.Experiments show that the proposed algorithm can improve the license plate recognition rate effectively.3)A training method about convolution neural network is proposed.The convolution neural network model has too many parameters and is prone to fall into local optimal solution.For this problem,a novel parameter initialization method is proposed.The K-means algorithm is used to train the first layer convolution kernels,and the redundant convolution kernels are removed.After the training process,the remaining convolution kernels are fixed.The experimental results show that the proposed parameter initialization method can effectively improve the license plate character detection rate and recognition rate.Then,the convolution neural network is optimized by adjusting the number of convolution kernels in convolution layers,the number of iterations and the ratio of positive and negative samples to get better result.Experiments have beenconducted on eight standard data sets from different sources.For license plate detection algorithm,the average accuracy rate is 95.64%.and for license plate recognition algorithm,the average recognition rate is 93.8%...