基于深度对抗网络和局部模糊探测的目标运动去模糊

Object motion deblurring via neural generative networks and local blur detection

目标运动场景去模糊问题是一个具有挑战性的病态逆问题,这是因为在动态场景中不同目标和背景区域可能会存在不同的模糊核.现有的基于能量优化的去模糊方法是将模糊图像分割成具有不同模糊度的多层图像,然后对不同的模糊层进行去模糊处理,然而其优化方案往往涉及迭代,耗时又烦琐.针对目标区域与背景区域可分离的模糊场景,结合传统的基于能量优化和基于深度学习方法的优点,提出一种基于深度对抗网络和局部模糊探测的目标运动场景去模糊模型,该模型由三个生成网络组成,用以建模潜在清晰图像、模糊核和模糊图像的权重变量.模型采用深度图像先验(Deep Image Prior,DIP)作为潜在清晰图像的正则化项,使用非对称跳跃连接自编码器生成潜在图像;采用全连接网络(Fully⁃Connected Network,FCN)生成模糊核.为了准确地获取模糊图像的分割结果,提出三条准则来指导权值变量网络结构的设计.实验结果表明,该方法同其他传统方法相比可以显著地提升重构性能,视觉效果更好.

Object motion deblurring is a challenging ill⁃conditioned inverse problem,because different blur kernels exist in object regions and background regions of the dynamic scenes due to movements in various directions and speed.Most existing energy optimization⁃based methods address this problem by segmenting the blurry images into multiple layers with different blurs,and then conducting deblurring process separately in different blur layers.However,they often involve iterative,time⁃intensive,and cumbersome optimization schemes,and the segmentation cannot be obtained accurately as a critical role in object motion deblurring.The purpose of this paper is to combine the advantages of deep Convolutional Neural Networks(CNNs)⁃based methods and conventional energy optimization⁃based methods.Aa self⁃tuning neural network is proposed for object motion deblurring which is composed of three generative networks to model the deep priors of clean image,blur kernel and the weight variables of object motion images respectively.In this network,Deep Image Prior(DIP)is used as the regularization of the latent clean image,and an asymmetric autoencoder with skip connections is adopted to generate the latent clean image.A Fully⁃Connected Network(FCN)is adopted to generate blur kernels.To obtain the segmentation of object motion images accurately,three rules are presented to design the specific generative network elaborately which produces the weight variables.Our preliminary experiments have been conducted for static scenes and object motion scenes deblurring,which show that the proposed method can achieve notable quantitative gains as well as more visually plausible deblurring results compared to state⁃of⁃the⁃art methods.