(A) study on motion deblurring using neural networks with cascaded enlarged receptive-field convolution units

Abstract

Motion blur occurs when an object or a camera moves fast especially under an environment with low light. Since the motion blur degrades the visual quality of images, much research has been made to solve the motion blur problem. Among them, many conventional methods have tried to directly estimate blur kernels and solve optimization problems using the blur kernels to restore the blurred images. However, these methods often failed to perform precise motion deblur and yielded artifacts such as halo effects and blocking artifacts. Recently, the motion deblur studies have intensively been made by using convolutional neural networks without estimating blur kernels directly. These methods have shown superior performance on motion deblur than the conventional methods. In this paper, we propose an Enlarged Receptive-Field Convolution (ERC) unit that can constitute effective and elaborate convolutional neural networks for motion deblur. Our proposed ERC unit is the cascade of a convolution layer with stride 2, a following convolution layer and a sub-pixel convolution layer with residual learning from its input to output. Our deep neural networks are constituted with cascade connection of multiple ERC units and can be trained in an end-to-end manner. The ERC unit allows for effective learning for motion deblur by convolution filters with enlarged receptive fields in lower-resolution feature maps. Our intensive experiments showed that our deep convolution networks with ERC units remarkably outperformed the state-of-the-art deep neural networks trained for motion deblur with maximum 2.26 dB higher in average PSNR.