Deep Idempotent Network for Efficient Single Image Blind Deblurring

Single image blind deblurring is highly ill-posed as neither the latent sharp nor blur kernel known. Even though considerable progress has been made, several major difficulties remain for deblurring, including trade-off between high-performance and real-time processing. Besides, we observe that current single networks cannot further improve or stabilize performance but significantly degrades when re-deblurring repeatedly applied. This implies limitation of these in modeling an ideal process. In this work, make two contributions to tackle above difficulties: (1) We introduce idempotent constraint into framework present a deep network achieve improved non-uniform with stable re-deblurring. (2) propose simple yet efficient lightweight encoder-decoder units recurrent structure can deblur images progressive residual fashion. Extensive experiments on synthetic realistic datasets prove superiority our proposed framework. Remarkably, nearly $6.5\times $ smaller notation="LaTeX">$6.4\times faster than state-of-the-art while achieving comparable high performance.