Research On Automatic Recognition Of Fuzzy License Plate Based On Deep Learning

At present,the clear license plate recognition algorithm has been mature,but there are few algorithms that can not recognize or identify difficult fuzzy license plates.The recognition rate of traditional license plate recognition algorithm is low or unrecognized.In view of this,a fuzzy license plate character recognition algorithm based on improved convolution neural network is proposed,which is combined with existing super-resolution algorithm and deblurring algorithm to recognize fuzzy license plates.The recognition rate is obviously improved compared with the traditional method.The research object of this paper is the segmentation of a single fuzzy license plate,which is a part of the study of character recognition,and does not involve regional license plate detection and segmentation.The main work and results are as follows:(1)The structure of a 11 layers convolution neural network for identifying the character of the fuzzy license plate is constructed.In order to improve the recognition rate of fuzzy license plate character,network model learning and optimization algorithms are studied: 1)The initialization of Xavier weight;2)Training to prevent overfitting,this paper added regular term in the objective function and increased the Dropout layer;3)Using the SGD optimization algorithm to optimize the network training.(2)Based on the factors of image degradation,this paper builded a fuzzy license plate sample set with different fuzzy degree.Subjectively,each character is divided into clear,semi clear and fuzzy three gears,that is to see clearly,seem to see clearly and not see clearly,forming from a clear to fuzzy gradient sample set,so that we can identify each type of license plate character with different definition.There are 26970 pictures in it.24273(90% of the total samples)sample images for training the convolutional neural network,the remaining 2697(10%of the total samples)samples as a test set,and trained 11 layers convolution neural network.The experimental results show that the 11 layers network for accurate identification of the training set rate reached 100%,and accurate recognition rate of test sets is 96.13%.(3)In this paper,a blind segmentation algorithm is proposed for blurred license plate with indistinct features,and each character is segmented accurately.In the image preprocessing stage,the perspective transformation method is used to correct the images.(4)From the angle of image sharpening,the deblurring algorithm and the superresolutionenhancement algorithm are introduced,and the experimental results are given.The results show that the effect of simple image processing is not ideal.(5)For the identification of fuzzy vehicle license plates,this paper analyzes two kinds of vehicle license plates(atomization non-atomization and)to do the experiment,and gives the recognition rate of 4 methods of 200 vehicle license plates including 1400 characters: 1)Calling training CNN network to recognize directly;2)Using super-resolution algorithm to process and then calling trained convolutional neural network to recognize;3)Using deblurring algorithm to process and then calling the trained convolutional neural network to recognize;4)Using traditional methods for character recognition.The experimental results show that the recognition rate of using super-resolution algorithm wih convolution neural network is 71.4%;The recognition rate of using deblurring algorithm to process with convolutional neural network is66.7%;Recognition rate of calling the trained CNN network directly is 57.1%;And the recognition rate of using traditional method is only 14.3%.In contrast,the recognition rate is obviously improved.