Deep convolutional neural network (CNN)-based restoration methods have recently received considerable progress in low level vision tasks, e.g., denoising, supperresolution, inpainting. A plain CNN, however, fails in image deblurring due to the severe pixel superposition caused by blur degradation. In this paper, we propose a novel concatenated residual CNN for image deblurring. Driven by minimum mean square error (MMSE)-based discriminative learning, the solution to image deblurring is interestingly unfolded into a series of iterative residual components, and is analyzed to demonstrate the contribution of iterative residual deconvolution (IRD). Furthermore, IRD motivates us to take one step forward to design CNN for image deblurring. Specifically, a residual CNN unit is employed to substitute the residual iteration, and is then concatenated and finally integrated, resulting in concatenated residual convolutional network (CRCNet). The experimental results demonstrate that CRCNet can not only achieve better quantitative metrics but also recover more visually plausible texture details.

本文提出了一种基于迭代残差卷积网络的图像去模糊算法，探讨了最小均方误差解决方案，将其解开成残差部分的序列，提出了一种由序列的残差卷积神经网络单元组成的连续的残差卷积网络。实验结果表明，提出的CRCNet不仅在定量指标上表现更好，而且在恢复视觉上具有更多的质感细节。

迭代残差图像反卷积