Research On License Plate Recognition Technology In Haze Weather

Combining modern information technology with traditional transportation system,intelligent transportation system demonstrates its safety,efficiency and environment-friendliness.License plate recognition is an important part of intelligent transportation system.Blurred traffic videos or images captured by imaging equipment in haze weather result in the loss of license plate details and low license plate recognition rate,thus bring difficulties to the implementation of intelligent transportation system.Aiming at this problem,this dissertation studies the recognition of blurred license plates captured in haze weather,and effectively solves the problem.The contents of research consist of four aspects,which are image defogging,license plate positioning,license plate character segmentation and license plate character recognition.Details of research are listed as follows:1.Aiming at the problem that the traditional defogging algorithm is not complete,a weighted fusion of dark channels defogging method is proposed based on atmospheric scattering model and dark channel prior defogging algorithm.In this method,the dark channel images are processed by histogram equalization and bilateral filter.Meanwhile,a novel method to calculate transmittance weight and light value in atmospheric scattering model are proposed.The final defogging images are obtained by weighted fusion result of the previous processed images and Single Scale Retinex algorithm.Experiments show that the images after defogging are more consistent with the fog-free images,and the haze images with dark color have better defogging effect.2.An image defogging method based on Generative Adversarial Networks(GAN)is proposed.In this method,a foggy image is put intothe generation model for training,and a new image is generated.It is up to the adversarial model to determine whether it is a clear image or a generated image.And then the parameters in the generation model are adjusted according to the maximum difference between the clear image and the generated image,until the adversarial model classifies the image generated by the generated model as a clear image.The defogging method is not based on the atmospheric scattering model,and has an effect more consistent with the mapping relationship between foggy images and clear images.3.A Reduced Yolov3-tiny Network(RYNet)is proposed to detect and locate license plates.RYNet improves the Yolov3-tiny target detection model by reducing the size of the model and using a multi-scale strategy.The experimental results demonstrate that RYNet has the advantages of low memory space cost and fast image detection speed while ensuring the detection accuracy.4.The slanted license plate is corrected by Radon algorithm,and then characters on the license plate are segmented with the vertical projection method.Finally,images of characters can be available by normalization.5.A vehicle license plate character recognition method based on Convolutional Neural Networks(CNN)is proposed.Making use of the difference between Chinese characters and non-Chinese characters(letters and numbers)on the license plate,a dual CNN structure are proposed to classify characters.This novel structure has the merit of high accuracy and easily-training.The proposed solution for license plate recognition in haze environment in this dissertation has the advantages of good defogging effect,high detection accuracy,fast detection speed and less memory space cost.And it has a good application prospect in intelligenttransportation systems.