红外深度图像的增强方法研究

红外深度图像的增强方法研究(论文13800字)
摘要：数字图像在采集、压缩、传输和渲染等各个阶段产生失真，这会严重影响图像质量。本论文涉及一种基于生成对抗网络的红外深度图像增强方法，其中包括图像的预处理和生成对抗网络模型的构建。生成对抗网络是由两个网络组成，分别是生成网络和判别网络；其中生成网络包括卷积网络单元、残差网络单元以及反卷积网络单元，判别网络由卷积网络单元构成。生成对抗网络通过基于MCL-3D图库的不同失真畸变的训练，形成用来增强图像的生成网络。通过测试，训练好的生成网络对大多数失真图像具有较明显的增强效果。
关键词：红外深度图像增强，生成对抗网路，卷积神经网络，残差网络
 
Research on Enhancement Method of Infrared Depth Image
Abstract：Digital images are distorted in various stages such as acquisition, compression, transmission and rendering, which will seriously affect the image quality. In this paper, a method of infrared depth image enhancement based on Generative Adversarial Networks is presented, which includes image preprocessing and construction of Generative Adversarial Networks model. Generative Adversarial Networks is composed of two networks, namely generation networkand discriminant network. Generation network includes convolution network unit, residual network unit and deconvolution network unit, and discriminant network is composed of convolution network unit. Generative Adversarial Networks is formed by training different distortion based on MCL-3D library to enhance image.Through testing, the trained generation network has obvious enhancement effect on most distorted images.
Key words: infrared depth image enhancement; Generative Adversarial Networks; convolution neural network;Residualnetworks
 

目录
1 绪论    1
1.1 研究背景与意义    1
1.2 研究现状    1
1.3 论文组织结构    3
2 生成对抗网络模型    4
2.1 图像的预处理    4
2.2 生成对抗网络    5
2.3卷积神经网络    6
2.4 残差网络    7
2.5 生成对抗网络的优化过程    7
2.5.1 判别网络的优化    7
2.5.2 生成网络的优化    8
2.6 生成网络模型    8
2.7 判别网络模型    9
2.8 本章小结    10
3生成对抗网络训练    11
3.1 生成网络的损失定义    11
3.2 判别网络的损失定义    11
3.3 Adam优化算法    11
3.4 判别网络与生成网络交替训练    12
3.5 训练过程与结果    12
3.5.1 高斯模糊    13
3.5.2 加性白噪声    13
3.5.3 采样模糊    14
3.5.4 JPEG和JPEG 2K压缩失真    14
3.5.5 传输损失    14
3.6 本章小结    16
4 实验结果    17
4.1 测试场景的参考图像    17
4.2 测试场景的六种畸变的失真深度图与增强图像    17
4.4本章小结    19
5 总结和展望    20
5.1总结    20
5.2展望    20
参考文献    22
致谢    24