Self‐supervised Perceptual Motion Deblurring using a Conditional Generative Neural Network Guided by Optical Flow

Jaihyun Koh

Abstract: The sample‐and‐hold characteristic of flat‐panel displays causes motion blur. Hardware compensation, such as frame‐rate doubling, is expensive, and existing methods of software compensation are slow. We propose a display motion deblurring network (DMDnet) which compensates for motion blur using a neural network trained on pairs of images with a synthetic random displacement between them. We assess the compensated images by convolving them with a kernel produced by a perceptual blur estimation algorithm, to simulate what is seen by a viewer. This technique is approximately 87 times faster than a state‐of‐the‐art optimization technique which produces an equivalent level of compensation; and its average PSNR is 29.70 dB, against 26.68 dB for the optimization.