Efficient License Plate Detection Algorithm In Natural Scenes

As an important part of intelligent transportation system,license plate recognition technology can provide efficient and practical management methods for intelligent transportation.The license plate recognition plays a critical role in the intelligent transportation system.In addition to being used in the toll gate system,the license plate recognition algorithm has also been used in the intelligent monitoring of various scenarios of urban traffic in recent years.Besides fixed-angle pictures,there are many license plate images taken on portable cameras such as drones and mobile phones.This type of images is usually of low quality,which is easy to cause problems such as blurring.This means that the algorithm has to handle more complex natural scenes.The algorithm needs to not only have good robustness,but also meet the requirements of real-time detection.Therefore,it is of great application value to study efficient license plate detection algorithms in natural scenes.The thesis focuses to study efficient license plate detection algorithms in non-ideal scenes.Due to the complex environment,variable angles and low detection efficiency,the current problems in current license plate detection technologies are addressed by deep learning based algorithms.The main innovations are as the follows.(1)In order to solve the problems of unbalanced license plate data and the unrealistic generation of license plates,a license plate generation algorithm based on the Generative Adversarial Networks(GAN)is studied.Using Cycle GAN to transform the artificially synthesized license plate data into the real data domain has alleviated the problem of the difference between the generated license plate and the real license plate distribution.By using Cycle GAN to convert a large number of artificially synthesized license plate data,the situation of unbalanced data volume during the training process based on deep convolutional network is solved to a certain extent,which brings a great performance improvement to the license plate character recognition algorithm.(2)The improved efficient license plate detection algorithm based on Spatial Transformer Networks(STN)is investigated.The use of spatial transformation networks can maintain spatial invariance and training features without the need for additional supervision information.By adding an STN module to the detection part of the license plate recognition network,it is possible to help the network to learn rotation invariance implicitly without additional supervision information training.The very small number of parameters of the STN module can also keep the network highly efficient.After experimental comparison,the addition of this module can effectively improve the accuracy of classifying the rotating and tilting license plate images,and improve the network’s robustness to the recognition of images with inclination angles.(3)The efficient algorithm of license plate detection in unconstrained scenarios is studied.By combining the Anchor-Based and Anchor-Free algorithms,an end-to-end efficient license plate detection network is constructed.The network includes a vehicle detection module,an attention module,and a license plate detection module.We use YOLOv3 as the basic network for vehicle detection,use shared features to reduce the amount of calculation,and use the local features to detect the key point position of the license plate.In addition,the attention module enhances the semantic features of the local area.Applying this module can reduce the interference of the unrelated area and then improve the detection ability.Finally,we obtain a effective real-time detection network.