Motion Deblurring Based on Convolutional Neural Network

Object motion blur results when the object in the scene moves during the recording of a single exposure, either due to too rapid movement or long exposure, leaving streaks of the moving object in the image and thus degrading its quality. In this paper, we present a method to solve the object motion blur problem in images with clear static background. Specifically, we propose an object motion deblurring algorithm that uses a convolutional neural network with six convolutional layers to deblur the image. Taking advantages of the strong ability of feature learning in convolutional neural networks, our method can remove the blurring effect of fast-moving object while keeping the clear background untouched. It is well known that neural networks are best driven by large data sets and more data means more benefits for training convolutional neural networks; therefore, we generated training set of 144,000 images and test set of 32,400 images. Through carefully designed training process, our model learned the ability of deblurring the blurred object while keeping the clear background. The experiment results show that our approach can generate superior results to a representative image deblurring algorithm that treats the same blurred object and clear background.