基于深度学习的运动模糊图像分割

【作者】 刘平；

【导师】 刘传才；

【作者基本信息】 南京理工大学 ， 计算机应用技术， 2019， 硕士

【摘要】 图像分割是计算机视觉领域的一个重要研究方向,在图像理解中扮演着重要角色。图像分割是指将图像的每一个像素分类,从而将图像分割成若干个具有相似性质的区域。在获取图像的过程中,大气、光照等环境因素或者相机抖动、对焦不准的人为因素,都会导致图像质量的退化,使图像变得模糊,传统分割方法不再满足这种退化图像的分割要求。随着深度学习技术的逐步深入,越来越多的研究人员开始使用卷积神经网络来解决计算机视觉领域的问题,涌现出许多优秀的模型。但是一种深度神经网络模型一般只解决一种问题,面对运动模糊图像的分割任务,本文采取串联两种深度神经网络模型的方式,来完成运动模糊图像分割的任务。本文的主要研究工作和创新点如下:首先,提出一种基于生成对抗网络的去运动模糊算法。此算法针对目前运动模糊图像复原过程中模糊核估计复杂、图像复原质量低等问题,将深度学习概念应用于图像复原领域,实现了端到端的去运动模糊算法的模型。在深度学习模型方面,考虑到清晰图像与模糊图像对应像素点比较接近,在模型的损失函数中将内容损失和Wasserstein距离相结合来约束图像内部特征,增强模型的有效性;在网络结构方面,根据残差网络和深度可分离卷积设计了残差可分离卷积模块作为生成网络模块,并加入了双线性插值缓解生成内容的马赛克,在判别模型中加入全局平均池化层来代替全连接层,对整个网络在结构上做正则化以避免过拟合,极大地减少了网络参数,对输入空间变换更具有稳定性;根据实验分析优化,获取了更好的复原率。最后通过实验验证该模型具有显著的复原效果。其次,将去运动模糊算法与最新的语义分割算法Deeplabv3+串联起来解决模糊图像分割问题。在深入研究诸如空洞卷积、空间金字塔池化、多孔空间金字塔池化等关键性技术的基础上,实现了 Deeplabv3+语义分割算法,并验证了复原图像的分割效果。随后利用马尔可夫过程随机生成模糊核的运动轨迹,通过将子像素插值应用于轨迹矢量来生成模糊核,从而人工合成VOC2012数据集对应的模糊数据集。将训练好的去运动模糊算法迁移到人工合成的模糊数据集上,再将新的训练好的去运动模糊算法与DeepLabv3+分割模型串联,对比实验结果验证了该思路的有效性。本文所提算法在Gopro数据集和VOC2012数据集上分别进行了大量的实验,并与现有算法进行对比,实验结果表明了所提算法具有较好的去运动模糊效果以及分割效果。更多还原

【Abstract】 Image segmentation is an important research direction in the field of computer vision and plays an important role in image understanding.Image segmentation refers to classifying each pixel of an image in order to segment the image into several regions of similar nature.In the process of images acquiring,environmental factors such as the atmosphere and illumination,or human factors such as camera shake and inaccurate focus,will lead to degradation of image quality and blurring of images.Traditional segmentation method couldn’t satisfy the segmentation requirements for such degraded images.With the gradual deepening of deep learning technology,more and more researchers have begun to use Convolutional Neural Network to solve problems in the field of computer vision,and many excellent models have emerged.However,a deep neural network model generally solves only one problem.In the face of the segmentation task of motion blurred images,two deep neural network model models are connected in series to complete the task of motion blurred image segmentation in this thesis.The main research work and innovations of this thesis are as follows:Firstly,a motion blurring algorithm based on Generative Adversarial Networks is proposed.This algorithm solves the problems of complex blur kernel estimation and low image restoration quality in the current motion blurring image restoration process.The deep learning concept is applied to the image restoration field,and the end-to-end motion deblurring algorithm model is realized.In terms of the deep learning model,considering that the sharp image is close to the corresponding pixel of the blurred image,the content loss and the Wasserstein distance are combined to constrain the internal features of the image in the loss function of the model,and the validity of the model is enhanced.In terms of network structure,the residual separable convolution module is designed as the Generative Model module according to the Residual Network and the Depthwise Separable Convolution.To ease the mosaic of the generated content,this thesis adds a bilinear interpolation to the generation network module.The global average pool layer is added to the Discriminative Model instead of the full connection layer,which can helps the structure of the whole network avoid over-fitting and greatly reduces the network parameters and makes the input space transformation more stable.According to the experimental analysis and optimization,a better recovery rate is obtained.Finally,the model is verified by experiments to have significant recovery effect.Secondly,the motion deblurring algorithm is combined with the latest semantic segmentation algorithm Deeplabv3+to solve the blur image segmentation problem.On the basis of in-depth research on key technologies such as Dilated Convolution,Spatial Pyramid Pooling and Atrous Spatial Pyramid Pooling,the Deeplabv3+semantic segmentation algorithm is implemented,and the segmentation effect of motion blurred image is verified.Then,the motion trajectory of the blur kernel is randomly generated by the Markov process,and the blur kernel is generated by applying the sub-pixel interpolation to the trajectory vector,thereby synthesizing the blur data set corresponding to the VOC2012 data set.The trained motion deblurring algorithm is migrated to the artificial blur data set,and the new trained motion deblurring algorithm is connected with the DeepLabv3+segmentation model.The experimental results verify the effectiveness of the idea.The proposed algorithm in this thesis has carried out a large number of experiments on GoPro dataset and VOC2012 dataset respectively,and compared with the existing algorithms,the experimental results has better motion deblurring effect and segmentation effect.更多还原