Research On License Plate Recognition System In Unstructured Scene Based On Deep Learning

Automobile is a very common means of transportation,and the number of automobiles is very large,which poses a great challenge to the management of automobiles by relevant departments.The license plate is the identity information of the car,and the accurate identification of the license plate plays a crucial role in the car management process.Therefore,it is of great engineering significance to develop a high-precision license plate recognition system.The existing license plate recognition system mainly aims at the license plate with small-angle deflection.In the unstructured scenario,the license plate will present a variety of tilt Angle including large-angle deflection.The accuracy of the existing license plate image recognition system with large-angle deflection cannot meet the needs of engineering applications.Based on this,in this paper,the license plate recognition system in the unstructured scene is deeply studied.A license plate correction algorithm based on image segmentation and traditional image processing is designed.An improved CRNN(Convolution Recurrent Neural Network)-based algorithm is proposed.A license plate recognition method is proposed.All the proposed algorithms are integrated and deployed in the cloud application.The research work of this paper has certain reference value and guiding significance for the application and development of the license plate recognition system in the future.The research work of this paper is as follows:(1)A license plate correction algorithm based on image segmentation and traditional image processing is designed for license plate images with large angle deflection in unstructured scenes.In this method,different levels of feature fusion are first considered,an adaptive weight feature fusion module AFFM(Feature Fusion module)is designed.A new image segmentation network AFF-Net is proposed.With similar accuracy,AFF-Net has faster inference speed than U-net and Deeplabv3+.After the license plate segmentation image is acquired by AFF-Net,a specific filter operator is adopted to obtain its boundary line.the license plate corner points are acquired through the fitted straight line of the edge line.the license plate image is rectified by perspective transformation.The experimental results show that the license plate correction algorithm proposed in this paper has better correction effect than the existing correction method IWPOD-Net.(2)For the corrected license plate image,an improved SC-CRNN model is designed for license plate image recognition.Efficient-b3 is used as the backbone of the model.Considering the different attention information at different locations,Sc SE(Spatial and Channel Squeeze &Excitation)is used as the attention module of the model.CCPD data is used for transfer learning to improve the recognition performance of CTPSD data.The experimental results show that the backbone using Efficient-b3 has higher recognition accuracy than Mobilenetv2,Resnet34,Xception and VGG-16 as the model.At the same time,the Sc SE module improves the recognition accuracy of the CRNN model more significantly than the CBAM module,and the model parameters are lower.(3)Based on the proposed license plate correction method and license plate recognition method,a license plate recognition system with large-angle deflection in unstructured scenarios is developed.In this system,yolov5 l is selected to realize the task of license plate detection,and then the license plate correction and recognition are carried out on this basis.Finally,the mode of instrument acquisition-cloud processing is adopted and deployed.The experimental results show that yolov5 l has higher detection accuracy than yolov3 in the license plate detection task.CCPD data is used for transfer learning,which can further improve the detection accuracy of the CTPSD data set of the detection method.On the CTPSD data set,the license plate recognition system proposed in this paper has higher recognition accuracy than the existing license plate recognition systems with correction modules.On the CPLD data,the license plate recognition system proposed in this paper achieves the best recognition accuracy.