A method for de-blurring license plate images based on deep learning

Abstract

The present invention proposes a method for removing motion blur of license plate images based on deep learning. The invention is divided into a data set preprocessing stage, a training stage and a testing stage. In the preprocessing stage of the data set, the license plate area in the image is determined, the license plate characters are segmented and the image size is normalized, and Gaussian noise is added to obtain the training set. In the training phase, the GAN is used to learn the image deblurring model, and the linear sum of the mean square error, gradient error, and discriminant error of the network restoration results is used as the network loss to alternately train the discriminator and the generator. In the test stage, the license plate characters are segmented and used as the input of the generator in turn, and the deblurred license plate image is obtained by combining the deblurred results according to the original order of the license plate characters. The model proposed by the invention effectively constrains the edge of the license plate image, thereby improving the quality of the license plate image to remove motion blur, and shortening the recovery time at the same time.

Description

technical field [0001] The invention belongs to the technical field of image processing, and relates to a method for deblurring a license plate image with motion blur, in particular to a method for deblurring a license plate image based on deep learning. Background technique [0002] Due to the influence of factors such as excessive vehicle speed, hardware limitations of the capture equipment, and lighting environment, the capture image has certain motion blur, which affects the accurate acquisition of the license plate number and has a negative impact on the management of urban traffic. License plate de-blurring refers to using the high-efficiency computing performance of the computer to restore the license plate image with motion blur through intelligent algorithms to obtain a clear license plate image. License plate recognition is an important part of traffic law enforcement, and the de-blurring of the license plate is beneficial to improve the recognition of the license ...

AI Technical Summary

Problems solved by technology

[0004] In recent years, scholars at home and abroad have conducted in-depth research and discussions on the estimation of motion blur kernels. Pan et al. proposed l 0 Regularization constraints estimate the image itself and its gradient in the middle, but when there are few zero-value pixels in the image, the sparse constraint of the image itself does not work, resulting in low restoration quality in license plate deblurring; Fang et al. proposed l 1 The regularization term constrains the sparsity of the image, which improves the quality of license plate deblurring, but there is some pixel distortion when restoring the image; Song et al. proposed to use the l 1 Sparse regularization constrains the image, and introduces a license plate recognition algorithm to optimize the deblurring effect, but the cyclic process of license plate deblurring and license plate recognition makes the entire license plate restoration process take too long

[0005] In recent years, with the rapid development of deep learning and the wide application in the field of computer vision, the problem of image de-blurring has been extensively studied. Svobode et al. proposed to use the convolutional neural network model to train the model of license plate image restoration. This method is limited by small The movement direction and length of the car within the range are not suitable for license plate de-blurring in complex situations; Nah et al. proposed to apply generative confrontation network to deal with motion blur in dynamic scenes, but the edge features of the restored image are not obvious, and motion still exists Vague

Images