Multi-scale Deformable Deblurring Kernel Prediction for Dynamic Scene Deblurring

Deblurring aims to restore clear images from blurred ones. Recently deep learning are widely used. Previous methods regard deblurring as dense prediction problems and rarely consider the inverse operation of blur. In this paper, we propose a multi-scale deformable deblurring kernel prediction network for dynamic scene deblurring which uses a coarse-to-fine method to predict the per-pixel deformable deblurring kernel and uses the fusion weight to integrate the latent images in different scales. Since the spatially variable blur scatters pixel information to surrounding sub-pixels and leads to the spatially and quantitively uneven distribution of latent pixel information, the per-pixel deformable deblurring kernel can adaptively select the sub-pixels and linearly combine them into the clean pixel for information aggregation. The multi-scale architecture helps the deformable deblurring kernel enlarge the reception field. The residual image is added to convolution result in each scale to supply refined edges when the kernel cannot cover the areas existing latent pixel information. Besides, we add local similarity loss to constrain deformable deblurring kernel's weight and offset which boosts the deblurring performance. Qualitative and quantitative experiments show that our method can produce competitive deblurring performance.