【作者】 王鹏；

【导师】 刘渭滨；

【作者基本信息】 北京交通大学 ， 信号与信息处理， 2021， 硕士

【摘要】 随着拍摄设备的普及,人们可以将日常生活中的场景以图像的形式记录下来。然而,在拍摄时由于相机抖动或物体运动,获取到的图像经常会出现运动模糊现象。因此,对模糊图像进行复原是十分必要的。模糊图像复原的过程,也被称为图像去模糊,这一研究方向是计算机视觉和图像处理领域的重要任务。最近几年,基于深度学习的“端到端”去模糊方法变得越来越受欢迎。本文通过搭建多个深度网络对运动模糊图像进行学习,从而复原出对应的清晰图像。主要研究内容可以分为如下三个部分:(1)为了解决复原模型难以获取全局信息的问题,本文提出了一个全局感知生成对抗网络。通过引入全局上下文模块来获取全局信息,可以使模型更容易处理图像模糊。另外,考虑到模糊图像缺少细节信息,本文还提出了一个空间细节增强模块来自适应地学习特征的空间信息,并对特定位置的信息进行增强,从而使图像细节更清晰。实验结果表明,该方法不仅具有较高的定量评估结果,而且可以很好地去除模糊并恢复图像细节。(2)为了利用不同尺度特征所包含的信息,本文提出了一个多尺度特征融合网络。首先,提出了一个跨尺度特征融合模块来对不同分辨率的特征进行融合,从而提升去模糊性能。另外,还提出了一个多尺度卷积块来获取不同感受野下的局部信息,通过对不同卷积核提取到的特征进行融合,模型可以更好地保留图像细节。最后,本文还在多个尺度上对图像进行重建,使模型可以更准确地预测复原图像。实验结果表明,该方法不仅可以获得更高的性能,而且具有更好的视觉感知效果。(3)为了有效地去除局部区域和全局图像中的模糊,本文结合前两部分工作提出了一个加权空间金字塔特征融合网络。首先,采用空间金字塔的形式将输入图像划分为不同区域,使模型分别学习到局部区域和全局图像的特征,并提出了一个加权特征融合模块来对这些特征中的信息进行融合。另外,考虑到模糊图像中的高频信息退化严重,本文还提出了一种高频增强模块,使图像中的细节更明显。最终的实验结果表明,该方法在定性和定量方面不仅优于前面的两种方法,也优于其他方法。更多还原

【Abstract】 With the popularity of shooting equipment,people can record scenes in daily life in the form of images.However,due to camera shake or object movement during shooting,the captured image often has motion blur.Therefore,it is very necessary to restore blurred images.The process of blurred image restoration,also known as image deblurring,is an important task in the field of computer vision and image processing.In recent years,the ”end-to-end” deblurring method based on deep learning has become more and more popular.In this paper,we build several deep networks to learn the motion blur image,so as to recover the corresponding sharp image.The main research contents can be divided into the following three parts:(1)To solve the problem that the restoration model is difficult to capture global information,a global awareness generative adversarial network is proposed in this paper.By introducing the global context block to capture the global information,the model can deal with the image blur more easily.In addition,considering the lack of detail information in blurred images,a spatial detail enhancement module is proposed to adaptively learn the spatial information of features,and enhance the information of specific locations,so as to make the image details sharper.Experimental results show that this method not only has high quantitative evaluation results,but also can remove the blur and restore the details well.(2)In order to take advantage of the information contained in the different scale features,a multi-scale feature fusion network is proposed in this paper.Firstly,a crossscale feature fusion module is proposed to fuse features at different resolutions,so as to improve the performance of deblurring.In addition,a multi-scale convolution block is proposed to obtain the local information of different receptive fields.By fusing the features extracted from different convolution kernels,the model can better preserve the image details.Finally,the image is reconstructed on multiple scales,so that the model can predict the restored image more accurately.The experimental results show that the method can not only obtain higher performance,but also have a better visual perception effect.(3)In order to effectively remove the blur in local regions and the global image,this paper proposes a weighted spatial pyramid feature fusion network combined with the first two parts.Firstly,the input image is divided into different regions in the form of a spatial pyramid,so that the model can learn the features of local regions and the global image respectively,and a weighted feature fusion module is proposed to fuse the information in these features.In addition,considering the serious degradation of high-frequency information in the blurred image,this paper also proposes a high-frequency enhancement module to make the details of the image more obvious.The final experimental results show that this method is not only better than the previous two methods in qualitative and quantitative aspects,but also better than other methods.更多还原