Research On Motion Blur Image Restoration Technology Based On Deep Learning

When photographing an object with a camera such as a digital product,the resulting image quality is blurred due to relative motion between the target object and the camera.The traditional methods have established the regular terms of the image restoration process through various methods.Although the ringing effect is suppressed to a certain extent and the noise is reduced,their restoration results are not very satisfactory.With the development of deep learning in recent years,the processed pictures can achieve a more ideal deblurring effect and process the motion-blurred images end-to-end.The latest multi-scale recurrent network model can better restore motion blur images,and the recurrent module effectively reduces the parameters in the network,making the model simpler and easier to train.However,the disadvantages are that the edge characteristics of the image restoration result are not obvious,and the high-frequency information is small.The reason is that it depends on minimizing the L₂ loss,which causes the restored image is too smooth for the optimal mean and produces artifacts.The final restoration results in visual perception do not meet our requirements,and they use large convolution kernels in the convolution layer,which leads to poor learning ability of features.To address the problem of the visual perception quality appears the edge features are not obvious and large-scale artifacts in the image motion deblurring.And In deep learning,when it training,there is a problem that more parameters are not easy to train due to increase the network depth,which is instability of its training.A multi-scale recurrent residual neural network was proposed.The model was improved on the basis of SRN-Deblur Net network.Firstly,in order to balance the evaluation metric(PSNR,SSIM)and visual perception quality and improve the high-frequency detail information in the image,making a multi-loss ensemble improvement.The perceptual loss and ₁L loss are combined.The perceptual loss is a simple ₂L loss.In model training,the training speed is fast and the convergence effect is good.It has high-frequency detailed information,and the ₁L loss function,no matter what input value is,it has a stable gradient,does not cause a gradient explosion problem,and has a more robust solution.Secondly,in the encoder-decoder network,a small convolution kernel stack is used in its convolution layer to deepen the network,making a smaller number of parameters and is easier to train.And better fitting the feature information of the picture.Finally,a certain optimization is performed in the residual network.In order to speed up the training speed of the entire network and improve the training generalization ability,the batch normalization layer(BN)after the second convolution layer in the residual block is retained.The BN layer after the convolutional layer has the following advantages:The BN layer can speed up the convergence of the network and prevent problems such as gradient explosions.Experimental results show that the proposed algorithm achieves good deblurring results.