Research Of Stack-based Multi-scale Recurrent Network For Image Deblurring

In recent years,the image deblurring task has attracted more and more researchers' attention.Many researchers are devoted to eliminating motion blur by using a “coarse-to-fine” architecture.And it has shown its superiority in eliminating motion blur which is caused by a simple relative displacement.But when using the“coarse-to-fine” architecture directly for the face image deblurring task,there still exist some problems:(1)Complex network structures make the model difficult to train and the large number of parameters results in expensive runtime.(2)Since the image details can't be restored well,quality of restored images will get worse,and deblurring visual appearance will be also insufficient.So in order to solve the above problems,in this paper,the ConvLSTM network(a special type of recurrent neural network)is employed.Since it contains a simpler network structure with “memory”,where the hidden state of recurrent modules can capture useful information and benefit restoration,it is helpful to recover images with more details.At the same time,in order to improve the quality of the restored image,by combining the “coarse-to-fine” and the stacked structure,we propose a new structure “Stack-based Scale-recurrent Network” for the face image deblurring task.And the residual module is used to optimize the network.Different from the other traditional methods with multiple losses,we have only one loss computation,which reduces the computation and improves the deblurring performance.Compared with other image deblurring methods,the results of this paper are obviously better than other methods,which shows that our method has higher deblurring performance.Based on that,in this paper,the proposed method is further optimized by exploring Adam algorithm.The research results of our paper have certain application value to the restoration of blurred face images,and it is expected to be a powerful way to solve the problem of face images motion blur.