Motion Blur License Plate Recognition Based On Convolutional Neural Network

In practical applications,license plate recognition may be caused by excessive speed,excessive exposure time,or blurring of the hand during shooting.As a result,the picture captured by the surveillance camera produces motion blur,which is even unrecognizable by the human eye.How to solve the difficulty caused by the blur in the captured image to the license plate recognition and restore the important semantic information of the license plate has great practical significance in practical application.Due to the large locality of the license plate data,there is still no unified database,and it is difficult to obtain fuzzy license plate data.In this theies,by collecting the license plate on the elevated Chengdu as the real fuzzy data,this theies attempts to use the simulated motion blur to simulate the motion blur license plate to make up for the lack of fuzzy data.In order to solve the problem of fuzzy license plate recognition,this theies designs the license plate location and based on the filtering method SSD recognition algorithm flow.For the license plate location problem,the SSD algorithm is used to build the recognition algorithm framework on the VGG16 basic network.Firstly,the license plate in the motion blurred picture is used to find the license plate outline and locate the license plate rectangular frame area.The influence of parameter variables in training and testing on the license plate data is discussed.The influence of different size,threshold and learning rate on the correct rate of the model is discussed.The optimal parameters are obtained by the control variables.The experimental results show that compared with the Faster R-CNN algorithm.The positioning model of the SSD algorithm performs better.For the deblurring preprocessing problem,non-blind convolution deblurring is performed with a certain blurring range.The noise is suppressed while maintaining the details,and the results of the two filtering methods for the license plate filtering are discussed.The SSD network structure with fusion filtering method is designed.The fusion method of this theies is designed in the pre-processing layer of the network structure.The picture is filtered and processed by subsequent segmentation and data enhancement structure.The data set of this theies has the characteristics of small size,uneven illumination and different inclination angles.For the license plate character recognition with the above characteristics,the SSD network of the fusion filtering mode is input.The character recognition type includes 65 characters in Chinese,English,and numbers.The recognition performance of the model was evaluated by observing the PR curve and MAP value of the test sample.At the same time,compared with the traditional algorithm,the end-to-end recognition algorithm also reduces the complexity of the algorithm and improves the recognition efficiency.Experiments show that using Wiener filtering combined with SSD network can effectively improve the correct rate of motion blur license plate recognition.In order to demonstrate the fuzzy recognition effect of the algorithm,the algorithm effect display system is designed.Through the system,the user uploads the picture that he needs to recognize,and displays the recognition result.The interaction is simple and easy to use.