基于对抗生成网络的图像修复应用

以下是资料介绍,如需要完整的请充值下载. 本资料已审核过,确保内容和网页里介绍一致.  
无需注册登录,支付后按照提示操作即可获取该资料.
资料介绍:

基于对抗生成网络的图像修复应用(任务书,开题报告,论文11600字)
摘    要
当今随着影像设备的普及,几乎人人都拥有了记录美好画面的设备,得益于这些设备,我们能够保留最美好的回忆,每个人都能创造出独一无二的作品。但是其中也不乏一些影响其成为一幅优秀作品的因素,比如一些遮挡,一些技术或设备缺陷造成的缺损。传统的图像修复[1]算法往往能够在某些方面能够有不错的表现,但是存在泛化能力比较差的缺点,无法作为一个图像修复的模型普遍适用于多种不同的缺失问题,往往都是由多个模型组合使用。
本文致力于应用深度生成对抗神经网络,寻找一种能够有较强适应性,能够作为单一的模型进行部署的方案。通过使用TensorFlow构建了DCGAN[2],在GPU上完成了DCGAN针对CelebA数据集的训练并且保存了完成训练的生成对抗神经网络的结构和训练参数,提取了网络结构中关于由100阶噪声生成图像的生成网络部分以及训练好的参数,并将其单独作为图像修复的核心,围绕生成生成网络构建一个修复优化问题。即通过缺损图像未缺失部分来使得生成器可以生成与原图非缺失部分最为相似的图像,从而将生成的图像的对应位置填充到原图的未缺失部分。简单的填充可能会有较明显的边缘,所以在填充的时候运用了一些泊松融合的方式。
实验表明,该算法能够在对噪声进行不错的填充,对较少的缺失能够有良好的修复效果,对较大的缺失能够做到内容上的正确填充。最后展现了一下修复效果,以及提出了一个假设,并对这个假设进行了一些讨论。
关键词:深度学习  对抗神经网络  图像修复  TensorFlow
 
Image completion application based on GAN
ABSTRACT
With the popularity of video equipment today, almost everyone has the equipment to record beautiful pictures. Thanks to these equipment, we can retain the best memories, and everyone can create unique works. But there are also some factors that prevent it from becoming a good work, such as defects caused by some technical. Traditional image restoration algorithms can often perform well in some aspects, but have the disadvantage of poor generalization ability. They cannot be used as an image restoration model and are generally applicable to many different missing problems, often composed of multiple The model is used in combination.
This paper is devoted to the application of deep generative adversarial networks to find a solution that can be more adaptive and can be deployed as a single model. By constructing DCGAN using TensorFlow, DCGAN is trained on the CelebA dataset on the GPU and the structure and training parameters of the neural network will be saved. The generation network will generate image from 100th order noise. Structure and trained parameters, and making them as the core of image inpainting, build a repair optimization problem around the generative adversarial networks. In another word, the missing image of the defect image enables the generator to generate an image that is most similar to the non-missing portion of the original image, thereby filling the corresponding position of the generated image into the unmissed portion of the original image. Simple fills may have more pronounced edges, so the Poisson Blend is used during the inpainting.
After experimentation, the algorithm can fill the random noise of the picture well, can have a good repair effect for fewer missing, and can correctly fill the content of large missing. At the end,it shows the repair effect, and puts forward a hypothesis, and some discussion on this hypothesis
Keywords: Deep Learning; GAN; Image Inpainting; TensorFlow

目录
摘    要    II
ABSTRACT    III
目录    IV
第一章  绪论    1
1.1 课题背景    1
1.2 课题研究现状    1
1.3 主要内容    2
第二章  文献综述    3
2.1生成对抗网络    3
2.2神经网络    4
2.2.1 传统神经网络    4
2.2.1 卷积神经网络    5
2.3深度卷积对抗生成网络    6
2.2.1 生成器G    7
2.2.2 判别器D    7
2.2.3 损失函数    8
2.4 SHANNON信息论    9
2.5 训练技巧    10
2.5.1 BatchNormalization    10
2.5.2 Leaky ReLU    11
2.5.3 动量梯度下降    12
第三章  模型设计    13
3.1 实验平台    13
3.1.1 实验环境的搭建    13
3.1.2 深度学习平台    13
3.2 修复算法设计    14
3.2.1深度对抗神经网络    14
3.2.2修复优化    16
3.2.3 第三部分,泊松融合    17
3.3 训练数据的预处理    18
第四章  模型应用与测试    19
4.1 训练可视化    19
4.1.1 信号变量与槽函数    20
4.2 效果展示    21
4.3 进一步探讨    23
参考文献    26
总  结    28
致  谢    29