图像盲复原是从一幅观测的模糊图像恢复出模糊核和清晰图像, 传统盲去卷积算法采用简化模型估计模糊核, 导致预测模糊核与真实值误差较大, 最终复原结果不理想。针对此问题提出一种基于改进残差模块的多尺度卷积神经网络模型, 采用端到端模式, 无需估计模糊核。提出了一种基于限制网络输入的改进Wasserstein GAN (WGAN), 增加了一层输入限制层, 能够限定参数初始值, 提高了网络收敛速度。设计了多重损失函数, 融合了基于多尺度网络的感知损失和基于条件式生成对抗网络的对抗损失。实验结果表明: 所提方法在定量和定性评价指标上优于已有的代表性方法, 并且运行速度比相近算法快了4倍。

The purpose of image blind deconvolution is to estimate the unknown blur kernel from an observed blurred image and recover the original sharp image. Conventional methods used simple models to estimate blur kernel, meaning mistakes were inevitable between estimated blur kernel and the real one. It would cause the final deblurred image unpredictable. A multi-scale convolutional neural network was presented based on the novel residual network. And it restored sharp images in an end-to-end manner without estimating blur kernel. Domain constraint layer was designed to the WGAN, it could restrict parameters initial values and accelerate convergence. A total loss function was designed including perception loss which was based on the multi-scale network and adversarial loss which was based on conditional GAN. Extensive experiments show the superiority of the proposed method over other representative methods in terms of quality and quantity. The method is 4 times faster than the similar methods.