Research On License Plate Super-Resolution And Recognition For The Compressed Surveillance Video

Recent years,our country is in the high-speed development of the economic and social period,the domestic security situation is grim.The state invested a lot of money to build safe city video surveillance project,and established a perfect video surveillance system.License plate is one of the most valuable observation targets in surveilance video.As monitor camera distance is far away and surveillance video is compressed,which led to surveillance video's quality is poor,and can't see the license plate through the naked eye clearly.To improve the clarity of the license plate in the compressed surveillance video and to identify the effective information in the license plate by using the super-resolution and license plate recognition.This technology can be applied to the areas of criminal investigation and traffic hit and run.The research background as well as significance of the license plate super-resolution and recognition for the compressed surveillance video in this paper.Its recent studying status is summarized,and several important techniques involved in super-resolution reconstruction and license plate recognition are listed and described.When the target motion amplitude is too large,directly using the optical flow method to register the license plate in compressed surveillance video will make the registration precision low.According to the problem,the face detection,tracking algorithm and registration algorithm based on the depth learning are proposed.In the view of the problem,an algorithm includes car face detection which based on deep learning,tracking and registration is proposed,this make sure the license plate is extracted and registered from coare fine.A matching traning library has been built for car face detection,Faster R-CNN algorithm is used to detect the car face area.Using the color characteristics of the license plate,the interest point is detected.Using the trajectory fitting and Kalman filter algorithm to optimize the license plate interest point coordinates.At last,the optical flow method is used for registration and the color features is utilized to extract license plate area finely.Combined with experimental results,the validity and robustness of the algorithm are analyzed,which proves the validity of the algorithm and the speed of registration is much faster than using the optical flow to registration directly.In this paper,the classical super-resolution and dequantization distortion convolution neural network models are analyzed,a unified processing convolution neural network model is proposed to solve the problem of super-resolution and quantization distortion effect in compressed surveillance video together.A matching character training library is made,which is trained by the proposed unified processing convolution neural network model.In order to improve the depth of the model the deep residual network is used,the gradient prior is introduced to use the characters'prior information.Experimental results prove that the proposed super-resolution and dequantization distortion unified processing network does contribute to the recovery of characters,and the effect is better than other advanced algorithms in this field.The proposed gradient priori is indeed beneficial to the reconstruction of characters.When the high compression rate is very high,the clarity of characters can't be improved by means of super-resolution and dequantization distortion.A multi-frame compression reduction license plate reconstruction and recognition algorithm which based on Generative Adversarial Networks is proposed,which make up for the shortcomings of characters can't be reconstructed well in the case of high compression ratio.This algorithm uses the super-resolution network model proposed in Chapter 3 as a generator network and also proposes a discriminant network suitable for the algorithm.Considering the car plate has fewer characters and categories,VGG networks are modified to be suitable for character recognition in this algorithm.Experiments show that the algorithm can indeed successfully reconstruct easy confusing characters to a certain extent,and the visual effect is better than the super-resolution and dequantization distortion model unified processing algorithm.The character recognition does make up for the shortcomings of poor character reconstruction under high compression condition.Experiments show that the recognition method does have a certain effect and can correctly identify a certain number of characters in the actual high compression license plate.