Deep pyramid generative adversarial network with local and nonlocal similarity features for natural motion image deblurring

Abstract

It is of great importance to capture long-range dependency in image deblurring based on deep learning. Existing methods often capture long-range dependency by a large receptive field, which contributes by deep stacks of local convolutional operations. Therefore, it restricts network representation ability and causes unpleasant artifacts of restored images. In this paper, we propose a deep pyramid generative adversarial network with local and nonlocal similarity features, called LNL-PGAN, for natural motion image deblurring. First, we propose a nonlocal feature block as an essential component of the pyramid generator for obtaining nonlocal similarity features at multiple levels. Second, we design a local feature block as another essential component to make a great balance between local and nonlocal similarity features. The local and nonlocal feature blocks capture meaningful short-range and long-range dependencies in the pyramid generator to increase network representation ability. Third, we design a multiscale generative adversarial loss to preserve edge details and facilitate sharp edge prediction of restored images, and we introduce a multistage training strategy to facilitate network training, which can further improve the quality of the restored image. Extensive experimental results demonstrate that the proposed method yields superior performance against state-of-the-art methods on natural motion image deblurring in terms of visual quality and objective index, and it can be used as a unified network for single and dynamic motion image deblurring.