Robust Single Image Deblurring Using Gyroscope Sensor

Abstract

Motion blur in an image is caused by the movement of the camera during exposure time; thus, awareness of the camera motion is a key factor in image deblurring algorithms. Among the various sensors that can be utilized while taking a picture in handheld devices, a gyroscope sensor, which measures the angular velocity, can help in estimating the camera motion. To achieve accurate and efficient single-image deblurring with a gyroscope sensor, we present a novel deep network with a flexible receptive field that is appropriate for training features related to the nature of the blur. Two specialized modules are sequentially placed in the proposed network to adaptively convert the shapes of the convolutional kernels. The first module directly transforms the kernel shape into the direction of the camera motion indicated by the gyroscope measurements. In the middle of the network, where the feature abstraction is sufficiently proceeded, the second module integrates features from the blurry image along with the information from the gyroscope to convert the kernel shape effectively, even when the gyroscope sensor is unreliable. Using a new gyro-image paired dataset, extensive experiments were conducted to show the effects of the reliability of the gyroscope measurements on the deblurring performance and to prove the effectiveness of our strategy.