Motion Deblurring by Fusing Multi-Stage Image Features

Motion blur, which is very common in daily life, will greatly affect the observation and analysis of images, not only reducing the quality of images but also hindering other research work in the field of images. Motion blur usually comes from the relative displacement between the target object and the camera during the exposure process. Coarse-to-Fine strategy has been widely used in various image deblurring studies for the architectural design of single-image deblurring networks. In this paper, we present MSDeblurNet, a new end-to-end deblurring network model based on the coarse-to-fine strategy by fusing multi-stage image features. We present multi-stage image feature fusion mode, which may augment shallow features in time while continually extracting deep features, allowing the network to acquire as many features of the entire image as possible and therefore achieve better deblurring outcomes. Finally, we carried experiments on the GoPro public dataset, and the results show that using multi-stage feature fusion improves deblurring effects in both subjective and objective evaluations; the proposed end-to-end deblurring model can also effectively remove ambiguity statistically and qualitatively.