Research On License Plate Recognition Algorithms In Complex Scenes

License plate recognition technology is one of the core technologies of intelligent transportation system,which can effectively improve the intelligent level of transportation and enhance traffic safety and efficiency.The license plate recognition technology can realize automatic monitoring and management of vehicles and reduce the occurrence of traffic violation accidents.In addition,it can also be applied to intelligent parking management,urban road toll and vehicle security and other fields,which is of great significance to urban management and safety production.However,the existing license plate recognition technology is limited by various factors such as light transformation,variable shooting angles and distances,and motion blur,resulting in low recognition accuracy in complex environments.In addition,the deep learning-based license plate recognition technology also suffers from the lack of training data and the imbalance of training samples.Therefore,this paper focuses on the study of license plate recognition methods in complex environments,aiming to improve the accuracy and robustness of license plate location and recognition,and help improve the operational efficiency and management level of intelligent transportation systems.The specific contents include:(1)To address the problems of difficult license plate location and low recognition accuracy caused by various factors such as imbalance of license plate samples,variable shooting angle and distance,and motion blur,a license plate location and recognition method in complex environments is proposed.Firstly,the training sample imbalance problem is solved by style migration of synthetic license plate images through Cyclic Generative Adversarial Network.Secondly,a MF-Rep Unet-based license plate localization method is proposed,which incorporates the improved VGG structure and Feature Pyramid Networks into the U-Net model to make up for the defects of small target license plates and tilted license plates that are easily missed in license plate localization.Finally,the attention mechanism assists Recurrent Neural Networks to improve the problem of semantic feature ambiguity due to image quality degradation,which improves the accuracy of character recognition.(2)For the problems of poor quality and low recognition accuracy of license plate images in dim and blurred scenes,a license plate recognition method based on low-light enhancement and super-resolution pre-processing is proposed.Firstly,a low-light license plate image enhancement method based on Retinex theory is proposed.The Multiscale-SE reflection network is designed for the noise problem,and the multi-branch convolution and channel attention mechanism are used to extract important image features,which solves the halo problem caused by excessive downsampling.Secondly,a fuzzy license plate restoration method based on super-resolution reconstruction is proposed.BLSTM is introduced to establish the pre and post association of character sequences,and L1 and Gradient Profile Prior loss functions are combined to enhance the character edge contour,effectively improving the accuracy of vehicle license plate recognition.Finally,multi-scale convolution is used to improve the feature extraction layer and improve the feature extraction ability of the model.(3)Based on the theories and algorithms mentioned above and combined with actual business needs,a license plate recognition system for complex environments is designed and developed,using Py Qt,Open CV and Pytorch deep learning frameworks as core technologies,providing a convenient and efficient solution for the application of license plate recognition systems.