Research And Implementation Of License Plate Recognition Algorithm In Complex Background

With the rapid development of the economy,the material needs and living standards of the people have increased year by year,which has led to an increase in traffic demand.As an important means of travel,motor vehicles have increased their holdings year by year.The traffic management pressure that comes with it is also increasing.The research on intelligent transportation system is imminent,and the license plate recognition system as its core part is the focus of research.So far,the license plate recognition technology researched at home and abroad has been mature in the scene of high-definition stillness,which can solve the problem of license plate recognition in general natural scenes,such as parking lot access system and highway toll system.However,in complex scenes,such as rain and fog,uneven illumination,etc.the existing technology can not be well applied,and the overall recognition rate is not high.Therefore,the license plate recognition technology in complex background still faces enormous challenges.In order to solve this problem,this paper combines the characteristics of complex scenes and applies the deep learning theory in recent years to design an end-to-end license plate recognition model based on deep learning.The model is composed of several convolutional layers and candidate regions.The pool of interest regions and the RNN network for license plate recognition.In the end-to-end transmission process,both the license plate location and the license plate character recognition can be completed at the same time,which simplifies the system workflow and improves the speed of license plate recognition.Firstly,the traditional license plate recognition technology is deeply studied.The traditional license plate recognition system consists of license plate location,character segmentation and character recognition.So far,the accuracy of each part can be very high,but the complex process and each algorithm The lack of close coordination leads to a low overall accuracy of the system,which is about 89.6%.In addition,most studies do not take into account the special circumstances of complex scenes.In order to solve this problem,the related knowledge of deep learning is studied,and the idea of end-toend network in deep learning can effectively solve the shortcomings of error accumulation in traditional license plate recognition systems.Then,according to the characteristics of complex scenes,some special image preprocessing is performed on the collected complex images,including image defogging processing,illumination compensation processing and de-motion blur processing,and simulation tests are performed on all preprocessing methods.The results show that the image preprocessing operation is helpful for image clarity and is ready for the next license plate recognition model.And then,using a more popular deep learning related technology in recent years,a unified deep neural model is designed,which can locate the license plate and identify the license plate characters in a forward transfer.The entire model can be trained endto-end,in addition to the parameters can be optimized during the training process.This approach uses an end-to-end network to solve both tasks in comparison to existing methods that use license plate detection and identification as two separate parts.It not only avoids the accumulation of intermediate errors,but also improves the accuracy of license plate recognition and speeds up the processing.In order to verify the accuracy of the model,a large amount of license plate data was collected.In the Windows 10 operating system,the PyCharm integrated development environment was combined with Keras 3.4,CUDA 9.0,CUDNN V7.1,Python 3.6 and other software technologies to train the designed model.And testing.In the model training,23562 images in the data set were manually labeled with image software,21206 of them were selected for model training,and the remaining 2356 images were verified.The test results show that the end-to-end network has improved the accuracy from 89.6% to 94.2% compared with the traditional license plate recognition system,and the data has been greatly improved.Finally,in order to facilitate the use of the model,all algorithms are packaged into deep learning license plate recognition system software on the QT platform.The test results of the software are consistent with the simulation results,in which the license plate recognition rate under normal background is 96.3%,and the license plate recognition rate under complex background is also 90%.After the recognition rate of the two cases is combined,the final overall recognition rate of the software is 94.2%,which is 4.6% higher than the traditional 89.6%.