Lightweight Rolling Shutter Image Restoration Network Based on Undistorted Flow

Binfeng Wang, Yunhao Zou, Zhijie Gao, Ying Fu

Abstract: Rolling shutter(RS) cameras are widely used in fields such as drone photography and robot navigation. However, when shooting a fast-moving target, the captured image may be distorted and blurred due to the feature of progressive image collection by the rs camera. In order to solve this problem, researchers have proposed a variety of methods, among which the methods based on deep learning perform best, but it still faces the challenges of poor restoration effect and high practical application cost. To address this challenge, we propose a novel lightweight rolling image restoration network, which can restore the global image at the intermediate moment from two consecutive rolling images. We use a lightweight encoder-decoder network to extract the bidirectional optical flow between rolling images. We further introduce the concept of time factor and undistorted flow, calculate the undistorted flow by multiplying the optical flow by the time factor. Then bilinear interpolation is performed through the undistorted flow to obtain the intermediate moment global image. Our method achieves the state-of-the-art results in several indicators on the RS image dataset Fastec-RS with only about 6% of that of existing methods.