Research On Vehicle License Plate Detection And Recognition Based On Deep Learning

The technology of automatic license plate detection and recognition in surveillance scene is the key technology in intelligent transportation system,and the emergence of new energy vehicle license plate brings new problems of license plate detection and recognition.What's more,due to factors such as the natural environment and monitoring equipment,the collected license plate images have both clear high-quality images and blurry low-quality images.Therefore,for different quality license plate images,it is necessary to design a detection algorithm that can correctly extract license plate information.This paper focuses on the problems of vehicle license plate detection and recognition under different quality surveillance videos,including the existing deficiencies of vehicle license plate detection and recognition technology,as well as the in-depth research and implementation of the traditional license plate and new energy license plate detection and identification problems under the monitoring scene.The main research contents are as follows:Firstly,a new energy license plate detection method based on video surveillance is proposed,and a new energy vehicle license plate dataset is established.The fully convolutional network is constructed to extract different "granularity" features of the license plate layer by layer from top to bottom,and then up-sampling to fuse the features of each layer from bottom to top,and then generate candidate frames based on different confidence according to the feature fusion map,and then different weight values are assigned to corresponding candidate boxes according to different confidence levels,and finally,a new energy license plate detection model with higher positioning rate in the monitoring scene is obtained.The experimental results show that the method of license plate location using fully convolutional network has a higher accuracy than other methods using single feature,and the undetected rate is significantly lower.At the same time,the accuracy of detection of traditional license plates is also greatly improved.Secondly,a one-stage identification of variable length license plate numbers is realized.The license plate recognition model in this paper does not require the preprocessing step of the license plate character segmentation.It uses convolution neural network(CNN)and LSTM(Long Short-Term Memory)-type recurrent neural network to build a four-layer convolutional neural network combined with long-term and short-term memory networks and CTC(Connectionist Temporal Classifier)sequence models,and it takes the entire character string in the license plate image as the recognition target,which can recognize both seven-digit and eight-digit variable length license plate character sequences,avoid the limitation of the license plate character segmentation,and at the same time,a higher-precision license plate recognition performance evaluation standard is proposed,which is more suitable for practical application requirements.Thirdly,a license plate detection and recognition application system based on the "vehicle pipeline" concept of the deep learning framework is realized.Vehicle detection is carried out by object detection framework YOLO3,and then vehicle images with complete license plates are input into the license plate detection module.The complete license plate images output by the license plate detection module are subsequently recognized without segmentation,and finally the recognition results of plate character sequence are obtained.At the mean time,the operation processing of image and video applications is designed for multi-threading processing of video and image.