NBD-Gap: Non-Blind Image Deblurring Without Clean Target Images

Image Compressive Sensing based on deep network has achieved a great breakthrough. However, there is still limited information on the details of the recovered structure. This paper proposes a Dual-domain Update and Double-group Optimization Network to solve the above-mentioned problem. Specifically, the Dual-domain Update Module establishes a new idea of iterative update by designing a pixel-feature domain, which effectively utilizes the richer information in the multi-dimension feature domain Double-group Optimization Module adopts the method of grouping denoising to reasonably utilize local information to improve the details of the recovered information. in addition, by introducing the inter-feature Iterative Recurrent Network to better retain and update the detailed information, the network has a more impressive reconstruction effect and appearance effect. Experimental results show that the network outperforms existing state-of-the-art methods and has better visual effects.