Deconvolution of Image Sequences with a Learning FFT-based Approach

In this work, we address the problem of fast and accurate non-blind motion deblurring of a sequence of frames. We propose an approach based on single image non-blind deconvolution methods by extending them from two spatial dimensions onto a 2D + time space. The resulted algorithms utilize recent advances of deep learning and heavily rely on learning the spatial and time correlations of frames from data which allows to perform the deblurring in an end-to-end fashion without the manual tuning of parameters. Developed algorithms were trained and tested on blurred video sequences, and as a result an improvement in restoration quality by around 1dB in terms of PSNR was achieved, comparing to a single-frame deconvolution. The proposed algorithms perform video restoration in a close to realtime speed, having potential for real applications in both portable consumer electronics and machine vision systems.