Stack-Based Scale-Recurrent Network for Image Deblurring.

Quite a few researches are devoted to eliminating motion blur using "coarse-to-fine" architecture. And it has shown its superiority in removing motion blur in some cases. However, there are still exiting problems: Complex network structure and large number of parameters, which makes the model difficult to train and result in expensive runtime. Poor quality and insufficient deblurring performance of images when simply using this architecture to remove motion blur. To solve the above problems, In this paper, we utilize the "coarse-to-fine" architecture and combine the benefits of stacked structure, proposing a new structure "Stack-based Scale-recurrent Network" (SSRN) for image deblurring task. Different from other traditional multi-scale methods with multiple losses evaluation, we only have one loss evaluation, and Scale-recurrent network (SRN) is employed in our model since it has simpler network structure and the results recovered through it are better than those via other networks. We choose the LFW face dataset for testing in our experiment. Due to the particularity of face images, in this way, we can verify whether our model has good deblurring performance for blurred face images. On the other hand, it can also laterally reflect whether the model can be applied to face recognition task (make the blurred face image clear so that it can be used for face recognition, the face recognition method will be described in the following chapters). We evaluate our model by means of calculating PSNR/SSIM value and accuracy of face recognition on a large blurring dataset with complex motion. As we expected, compared with other deblurring methods, both the PSNR/SSIM value and face recognition accuracy of deblurred image obtained by our model have improved significantly, which confirms the superiority of our proposed model in image deblurring task. At the end of this article, we display some deblurred images to show the high deblurring performance of our model in deblurring task.