Deep Dynamic Scene Deblurring From Optical Flow

Deblurring can not only provide visually more pleasant pictures and make photography convenient, but also improve the performance of objection detection as well tracking. However, removing dynamic scene blur from images is a non-trivial task it difficult to model non-uniform mathematically. Several methods first use single or multiple estimate optical flow (which treated an approximation kernels) then adopt non-blind deblurring algorithms reconstruct sharp images. these cannot be trained in end-to-end manner are usually computationally expensive. In this paper, we explore remove by using multi-scale spatially variant recurrent neural network (RNN). We utilize FlowNets two consecutive different scales. The estimated provides RNN weights scales so that better help RNNs feature spaces. Finally, develop convolutional (CNN) restore deblurred features. Both quantitatively qualitatively evaluations on benchmark datasets demonstrate proposed method performs favorably against state-of-the-art terms accuracy, speed, size.