Research On License Plate Detection Algorithms In Complex Traffic Scene

As a fundamental element License Plate Detection(LPD)plays an important role of modern Intelligence vehicle system.LPD use optical imaging devices capturing images from traffic environments,after that the other parts of LPD automatically locate the license plates in above captured images and acquire exact coordinates,finally segment license plates’ areas by using computer vision,machine learning and other related technologies.After long-term development,LPD has been widely used in various fields of Intelligent Transportation,such as: Parking Lot Charging System,Traffic Road Flow Monitoring,Traffic Violation Vehicle Capture,Vehicle Intelligent Tracking and Location.At present,there are many theoretical research results and commercial systems for LPD,it seems that,the license plate detection technology is quite perfect and satisfactory results can be obtained by existing solutions.But in fact,such technology has many preconditions and limitations.For example,the commonly used Parking Lot Charging System should be installed under good acquisition environment with illumination,shooting angle,photographic stability in ideal state.The detection effect of the existing technology will be greatly reduced,once these preconditions be removed.So,study License Plate Detection technology in complex scenes has great practical value.This paper described our works of research on LPD in complex non-ideal environment of daily shooting environments.We do research by combing Deep Learning technology,to solve the problems of non-ideal shooting environment,multiangle shooting,deblurring restoration in LPD.The main innovations of our research are as follows:(1)We propose A license plate detection algorithm for non-ideal illumination scenes.Based on deep learning image semantics segmentation technology,it extracts robust image features and generates the segmentation map of the saliency region of license plate.Our network replaced fully connecting layers in common Convolutional Neural Network by convolution layers;We designed 4 dilation convolution branches,each of them can extract visual features with different receptive fields.We fuse the features of different branches and different layers of networks together,the segmented graph of the same size as the original graph is directly generated.The robust visual features extracted automatically by CNN and the automatic nonlinear classification effect generated by the activation layers in the CNN,the segmentation graph generated by the network has quite high detection capability.Compared with the different algorithm on two open license plate datasets,the proposed algorithm has higher detection rate in non-ideal shooting environment.(2)We propose a lightweight segmentation network for real-time license plate detection tasks.The goal of the network is to provide high speed image segmentation and control the precision reduction to an acceptable range to solve the problem of inefficient operation of ordinary segmentation algorithm.The main idea is reducing the amount of calculation and resource consumption by cutting down the numbers of each layers in network.The decrease of resource consumption will inevitably lead to the loss of accuracy,the network uses a parallel architecture to remedy this problem.Image features of different receptive fields processed by different branches,only one branch that consumes the least passes the complete convolution layer.The characteristics of different receptive fields can be obtained by overlapping several branches at the end of the network.The experiment result shows that a balance of efficiency and precision is achieved.(3)We propose an irregular shape license plate quadrangular location algorithm in arbitrary capture scenes.The proposed algorithm consists of a corners prediction network,a corner matching algorithm and a corner regression network.The proposed algorithm is suitable for dealing with license plate images with rotation,perspective distortion from non-frontal shooting.We adopt segmentation network outputting 4 corner heatmaps of license plates with 3 scales.The predicted corners are normally distributed in gray scale.After smooth superposition,peak values of processed heatmaps are coordinate points.For those images that contain multiple license plates,the number of predicted coordinate points are multiple of 4.We group the coordinate points by computing the proportion of interconnections crossing each other.The corner regression network finetune the above-mentioned coordinate points by computing relative distance.The proposed algorithm solved the problem that slide window cannot detect irregular quadrilateral in license plate detection.(4)We propose a deblurring algorithm for license plate image in unstable capture scenes effectively improve detection accuracy.The algorithm is trained by Generative Adversarial Network.In antagonistic training,Wasserstein distance with penalty term is used as loss function in discriminant model.The loss of discriminant model is limited to a controllable range,so that the model can accelerate convergence.The generate model adopts multiple upsample structure,which effectively utilizes different magnification calculations to generate the details of the deblurred image.The loss function of the generated model uses two structures: the mean square error of the feature space after feature extraction,and the improved mean square error of the generated image and the clear image in the intuitive pixel space.These two make it possible to compare the differences between the two spaces in the generated model,which helps to generate more realistic and clear images.The experiment result shows that license plate detection on defuzzied images has greatly improved the detection rate and recognition accuracy compared with the original image.