Blind Deblurring Algorithm For Motion Blur Of Single Picture

As an important information transmission medium,the image may be blurred and degraded due to the interference of various factors during the acquisition process,which brings difficulties to the subsequent processing of the image.Image motion blur is a kind of blur caused by the relative displacement of the object and the camera during shooting.This article mainly studies the method of image motion blur restoration.With the continuous development of deep convolutional neural networks,deep convolutional neural networks have achieved significant results in the field of deblurring,but there are still many challenges in the field of deblurring.The design of an effective deblurring network,the selection of a suitable loss function,and how to recover more image details have become urgent problems to be solved in the field of deblurring.The paper makes an in-depth study on the blind restoration algorithm of single-picture motion blur,focusing on the end-to-end blind deblurring method based on deep convolutional neural network.The thesis first studied the feature extraction module in the design of deblur network,followed the traditional deblur method from coarse to fine,and took the scale-recurrent convolution deblur network as the benchmark model,using the ordinary residual module and the multi-scale residual module And the improved multi-scale residual module is used as the feature extraction module of the network.By comparing the network deblurring results,it is verified that the multi-scale residual module and the improved multi-scale residual module have stronger feature extraction capabilities,and the restored image has more rich details.The improved multi-scale residual module has more advantages than the multi-scale residual module.It not only can realize the fusion and multiplexing of image feature information of different scales,but also greatly reduces the amount of parameters.By analyzing the results of image deblurring,it is found that using a data set with different degrees of blurring to train a deblurring network,the network has a large difference in the restoration effect of images with different degrees of blurring.The PSNR/SSIM of severely blurred images is significantly lower than that of slightly blurred images.Therefore,the motion blur evaluation factor is proposed to classify the blurred image into large-scale blur and small-scale blur.According to the classified blur dataset,the deblurring network is scaled to obtain the corresponding large-scale deblurring network and small-scale deblurring network.The deblurring results of different models at the same scale are merged to obtain the final deblurring result.The experimental results verify the effectiveness of the sub-scale deblurring method and improve the deblurring effect.