【作者】 李杰；

【导师】 周越；

【作者基本信息】 上海交通大学 ， 控制科学与工程， 2019， 硕士

【摘要】 近年来,降质图像复原一直是数字图像处理领域的重点研究对象,其中多因素复合下的降质图像复原问题更是这一领域的研究热点和难点。本文首先针对单因素降质图像复原问题提出了基于深度卷积神经网络的复原算法,随后将其扩展为解决噪声、模糊和降采样三种降质因素复合问题的复原系统,最后本文还提出了基于图像感兴趣区域的质量评价算法用于评价复原图像的质量。在单因素降质图像复原方面,本文以稠密卷积块为基础结构提出了一个端到端的深度卷积神经网络模型用于学习降质图像到清晰图像的映射关系。针对深度神经网络训练过程中出现的梯度消失问题,引入了局部跨连接和全局跨连接策略来加强特征信息与梯度信息的传递,同时使得复原网络以残差学习的方式专注于复原降质图像缺失的信息,保证了复原网络的高效性。在多因素降质图像复原方面,本文以单因素降质图像复原网络模型为基础,结合降质图像先验估计方法提出了解决复合降质图像复原问题的可行方案。其首先通过降质图像先验估计算法得到噪声先验和模糊核先验信息,然后使用维度变换策略对先验信息进行降维并扩展成与降质图像尺寸一致的先验图像,最后将其与降质图像一并作为复原网络的输入样本并训练得到复原网络模型,达到图像复原的目的。在降质图像质量评价上,本文结合了图像显著性和边缘强度来提取感兴趣区域,并以该区域为指导分别对降质图像的结构分量图像和纹理分量图像进行子图像块划分和筛选,随后使用本文提出的双端卷积神经网络模型对输入的结构分量子图像块和纹理分量子图像块进行质量评分,最后将所有子图像块的得分加权得到降质图像的质量评分。在公开数据集上的大量实验结果论证了本文提出的单因素降质图像复原网络模型的合理性和有效性,同时通过在自拍并合成的SJTU数据集上的大量实验论证了本文提出的多因素降质图像复原系统的可行性。最后在公开的图像质量评价数据集上通过实验论证了质量评价算法的准确性,并通过对复原图像进行质量评价验证了质量评价算法的有效性。更多还原

【Abstract】 In recent years,image restoration has always been the focus of digital image processing.Among them,the research on multiple degraded image restoration is a hot and difficult task.In this thesis,we firstly propose a deep convolutional neural network to solve the image restoration problem under a single degradation factor,and then extend it to a multiple degraded image restoration system for denoising,deblurring and image super-resolution.Finally,an image quality assessment algorithm based on regions of interest is proposed to evaluate the quality of reconstructed images.For single factor degraded image restoration,an end-to-end deep convolutional neural network based on dense blocks is used to learn the map function from degraded images to sharp images.To handle the gradient-vanishing problem during the training process,local and global connections are proposed,which improve the flow of gradient and features.At the same time,local and global connections make the restoration network focusing on restoring the missing information of degraded images in the way of residual learning,which ensures the high efficiency of the restoration network.For multiple degraded image restoration,we propose a feasible solution based on the single factor image restoration model and image priors to solve the multiple degraded image restoration problem.The prior information of noise level and blur kernel is obtained through prior estimation methods.Then,priors are reduced and conversed to prior images which have the same size as degraded images by dimension transformation.Finally,all prior images and degraded images are combined to train a image restoration model.For image quality assessment,this thesis firstly combines image saliency and edge intensity to calculate regions of interest and then crops image patches from structure components and texture components of degraded images.After that,we propose a dual-network to evaluate qualities of all those patches.Finally,all the quality scores are weighted to obtain the score of the whole degraded image.Abundant experiments on public datasets show that the proposed model is reasonable and effective.In addition,a series of experiments on the SJTU dataset demonstrate the feasibility of the proposed multiple degraded image restoration system.Finally,we also demonstrate the accuracy of our image quality assessment algorithm by experimental results on public datasets and show the effectiveness for restored images’ quality assessment.更多还原