基于区域选择网络的图像复原及其在计算成像中的应用

Image restoring method based on region selection network and its application in computational imaging

本文针对波前编码成像,单透镜计算成像等领域的全局一致模糊复原背景需求,提出了一种高效的基于区域选择网络的图像去模糊方法。与传统方法通过构建目标函数及各类先验信息实现模糊图像清晰化过程不同,本文方法则基于深度学习与传统方法的结合。传统方法负责图像复原的主体流程,深度学习方法则负责对传统方法中的关键步骤模糊核求取区域选择进行干预。基于深度学习的深度二元分类网络能够自动在全局图像中剔除平坦过曝、短小纹理等区域,并选取最优的用于模糊核求取的图块区域。传统复原方法则以此为基础实现模糊核求取,非盲图像复原及图像清晰化处理过程。实验结果表明:本文的复原方法能够实现良好的复原效果,纹理清晰,稳定可靠;所提出的区域选择网络能够在降低计算复杂度的同时,有效提升模糊核的估计准确度,进而提升图像清晰化的复原效果。在同等条件下,所提出的深度二元分类网络在误差率限定在1.5时,复原成功率较比现有方法提升了2.1%,同时复原图像的平均峰值信噪比较比现有方法提高了0.5 dB。

Computational imaging is a new interdisciplinary subject that has gained widespread research attention in recent years.However,its efficiency and recovery effect restrict its development in engineering applications.In this paper,an efficient image deblurring method based on a region selection network is proposed to tackle the restoration task in the fields of wavefront coding imaging and single lens computational imaging.In contrast to traditional image restoration methods,which usually involve construction of an objective function and addition of reasonable image priors to restore blurry images,the proposed method is based on a combination of a deep learning method and a traditional restoration algorithm.The traditional method is used for the main image restoration process,while the deep learning method is used to intervene in the kernel estimation region selection.The deep learning method involves constructing and training a deep binary classification network,which can automatically eliminate the flat overexposure,short texture,and other areas in the global image,and select the most suitable block area for kernel estimation.On this basis,traditional restoration methods perform kernel estimation,non-blind image restoration,and image enhancement processing.The experimental results show that the proposed method can achieve a good and stable restoration effect,that the proposed region selection method can reduce the computational complexity,and that the point spread function can be estimated well.When the error rate is limited to 1.5,the restoration success rate is improved by at least 2.1%,and the average peak signal-to-noise ratio(PSNR)is increased by at least 0.5 dB.