Deep Learning-Based Deblur Using Gyroscope Data

Global scene blur of an image is caused by the movement of a camera during the exposure time, thus camera motion is one of the key factors in identifying blur artifacts. Among various sensors that can be utilized while taking a picture in a mobile environment, a gyroscope sensor is useful for deblurring tasks, because rotational velocities measured from the gyroscope can be processed to estimate the camera motion. For enhancement of deblurring performance in a mobile environment with an additional gyroscope sensor, this paper presents a method of using gyroscope data in a deep learning-based deblur network. The proposed method calculates homography matrices from the gyroscope data and warps the input image to imitate the camera motion. The stack of warped images aligns blur artifacts caused by camera movement channel-wise, and the network is trained to use image features related to the direction and magnitude of the blur. The proposed method shows superior performance in cases of both subtle and severe blur conditions compared to conventional gyroscope-aided deblurring network.