Research And Implementation Of Super-Resolution Reconstruction Of Blur License Plate Image In Video Surveillance

License plate image is an important information source in surveillance video, but there are many kinds of degradation phenomena affecting the identification, such as low resolution, blur, noise and so on. Therefore, it is of great significance to reconstruct the license plate image effectively. This paper analyzes the specific features of the license plate images in surveillance scene, and conducts researches on the theoretical basis and key technologies of no reference image quality evaluation, license plate image preprocessing, image classification, and image super-resolution reconstruction. In this paper, we propose the main idea of reconstruction after recognition, and construct a novel license plate image reconstruction algorithm. The main research contents are as follows:1. This paper presents a license plate image quality evaluation method based on subjective classification features. By analyzing the quality of license plate image from several subjective classification features, this method gives a judgment and analyses the reasons when the quality of the target image is too poor to be reconstrcted.2. This paper proposes a geometric correction method suit to license plate images of different quality. Through the study of license plate image geometric correction methods and their applications, we adopt the method using Hough transform and Radon transform when establish the database of license plate image, and propose using a new geometric correction mothod based on perspective transformation when reconstruct a low resolution image.3. To carry out effective character recognition for license plate images, we adopt two character recognition methods. One is based on template matching and the other is based on discriminative dictionary learning. The method based on template matching uses a database lager than normal method to increase the stability of geometric position offset, and the recognition result is close to the human eye. The method based on discriminative dictionary learning constructs a discriminative dictionary through the constraints of the difference between the encoding vectors of different classes, and obtains better results in low quality situation. Meanwhile, a novel recognition algorithm based on the block of character is proposed after that. This algorithm considers the local similarity and global connectivity of characters to enhance the stability.4. In order to use the existing recognition results to carry out effective license plate image reconstruction, a super resolution reconstruction method based on sparse representation is adopted. After the super-resolution reconstrction of every character, we consider the features of the license plate and use gray scale histogram processing to enhance the image quality. In the end, all the characters are combined as a whole image.