Iterative Dual CNNs for Image Deblurring

Image deblurring attracts research attention in the field of image processing and computer vision. Traditional methods based on statistical prior largely depend selected type, which limits their restoring ability. Moreover, constructed model is difficult to solve, operation comparatively complicated. Meanwhile, deep learning has become a hotspot various fields recent years. End-to-end convolutional neural networks (CNNs) can learn pixel mapping relationships between degraded images clear images. In addition, they also obtain result effectively eliminating spatial variable blurring. However, conventional CNNs have some disadvantages generalization ability details restored image. Therefore, this paper presents an iterative dual CNN called IDC for deblurring, where task divided into two sub-networks: detail restoration. The sub-network adopts U-Net structure semantical structural features image, restoration utilizes shallow wide without downsampling, only texture are extracted. Finally, deblurred multiscale strategy that improves robustness precision model. experimental results showed proposed method excellent effect real blurred dataset suitable application scenes.