乌有摘要
变量的定义图像去雾技术在很多领域，诸如国防、交通、航空、天文研究等领域具有广泛的应用。现实中在雾、霾等天气下，由于受到大气中悬浮颗粒、汽水等影响，摄像头采集到的图像的质量很差，很大的限制了图像采集系统的应用效果。如何快速、有效地消除有雾天气对采集图像和视频质量的影响，从而提高图像清晰度与分辨度，这是图像增强领域的一个研究热点和难点。
基于深度学习的去雾模型相对传统的去雾模型，存在以下优点：首先，通过对海量的带雾和去雾图片样本进行不断学习，不断迭代更新学习参数，最终收敛的模型可以学习到更好的去雾特征。其次，针对不同场景下，基于深度学习的去雾模型具有更好的鲁棒性；再次，基于深度学习的去雾模型可以通过适当控制模型的参数、优化模型的卷积操作，使模型轻量级化。最后，通过有效地减少模型的冗余参数，可以加快模型去雾的处理速度，从而实现对带雾图像的实时去雾处理。
针对图像去雾算法要求高效、快速，应用终端设备要求可移动、大小、功耗有限等的应用场景，本文提出了一种基于深度学习的轻量级图像去雾算法模型LMDehaze-Net(Lightweight Mobile Dehazing Network)，并在移动端NVIDIA® Jetson™ TX1实现了其相关基础应用。本文主要工作内容包括：
（1）对传统去雾算法和最新端到端去雾算法进行了深入地研究，认真分析和总结了它们算法的特点。其次，针对模型的轻量化，对多种表现优秀的轻量级网络的卷积优化操作进行了详细地总结。
（2）针对已有去雾算法参数量大、图像去雾速度慢的问题，提出了一种基于深度学习的轻量级图像去雾算法模型LMDehaze-Net。基于公开数据集RESIDE和NYU2，将LMDehaze-Net模型与传统算法模型、最新端到端算法等模型进行了多方面的实验效果对比。实验结果表明LMDehaze-Net能够高效、快速地解决图像去雾问题。
（3）针对一般的深度学习平台存在高性能显卡体积大、价格昂贵、模型部署困难等问题，本文采用了先在实验室GPU集群上训练好LMDehaze-Net模型，然后将训练好的模型部署到内嵌信用卡大小GPU的Jetson™TX1硬件平台上的方法，并在Jetson™TX1上实现了LMDehaze-Net算法模型的相关基础应用。
关键词：去雾算法；轻量级；深度学习；LMDehaze-Net；Jetson™ TX1
Abstract
Image dehazing technology is ud in many fields widely, such as national defen, transportation, aviation and astronomical rearch. In reality, as affected by suspended particle, vapour, etc. in the atmosphere, the quality of the image obtained by the camera is very poor under hazy weather, which creates lots of limits to the application effect of image acquisition system. It is a hot and tough rearch topical in the field of image enhancement that how to eliminate the hazy weather to the infl
uence of image or video quality, and improve image clarity and resolution quickly and effectively.
Compared with the traditional dehazing models, the models bad on Deep Learning has the following advantages: Firstly, The final convergent deep learning dehazing model will catch the better dehazing features from the mass coupled pictures that contain lots of hazy-images and their corresponding dehazy images.Secondly, for different scenarios, the models bad on Deep Learning has better robustness. Thirdly, the model can be lightweight by controlling model parameters and optimizing model convolution operation. Finally, the speed of model dehazing can be accelerated, and the real-time dehaze for the hazy images can be realized by reducing the redundant parameters of the model effectively.
羊肉补什么To solve the problems that require the dehazing algorithm is efficient、rapid, and terminal devices require mobility, limited size and limited power consumption. In this paper, we propo a lightweight image dehazy algorithm model called LMDehazeNet (Lightweight Mobile Dehazing Network)bad on Deep Learning, and the basic applications of this algorithm have been implemented on th e mobile terminal NVIDIA® Jetson™ TX1.The main work of the paper includes the following:
(1) In this paper, the traditional dehazy algorithms and the latest end-to-end dehazy algorithms are st罗伯特卡茨
udied deeply. In addition, their characteristics are analyzed and summarized in detail. For aspect of the lightweight model, the convolution optimization operations of many excellent lightweight networks are summarized.
(2) To overcome the problem of existing dehazy algorithms, which large amount of parameters in the model slows the speed of dehaze. A Deep Learning lightweight image dehazy algorithm model called LMDehaze-Net is propod. Bad on open datats,including RESIDE and NYU2, LMDehaze-Net algorithm model is compared with other traditional algorithm model and the latest end-to-end algorithm model though experiments.The results show that our algorithm can solve the problem of image dehaze efficiently and quickly.
(3) To avoid the problems of large size, high price and difficult deployment of high
performance graphics cards in general deep learning platforms, we decide that the LMDehaze-Net model should been trained on the laboratory GPU clusters, and then deployed the trained model on the Jetson TX1 hardware platform which embedded credit-card size GPU. The related basic applications of LMDehaze-Net algorithm model is realized on Jetson TX1.
Key words:Dehazing Algorithm; Lightweight; Deep Learning; LMDehaze-Net; Jetson™ TX1
目录
摘要........................................................................................................................................... I Abstract ..................................................................................................................................... II 第一章绪论 (1)
1.1 课题研究背景及意义 (1)
1.2 国内外研究现状 (2)
差多音字1.2.1 基于物理模型的图像去雾算法 (2)
1.2.2 基于图像特征差异的图像去雾算法 (3)
1.2.3 基于端到端深度学习的图像去雾算法 (4)
1.3 本文研究的内容 (5)
1.4 本文组织结构 (5)
第二章轻量级去雾模型设计基础 (7)
2.1 概述 (7)
2.2 经典大气光散色物理模型 (7)
2.3 传统的去雾算法 (9)
2.4 基于深度学习的去雾算法 (11)
2.5 轻量级网络 (14)
2.5.1 Inception中的卷积结构 (15)
2.5.2 MobileNet中的深度可分离卷积 (16)
2.5.2 ShuffleNet中的逐点群卷积和通道混洗 (17)
2.5.3 EffNet中的空间可分离卷积 (18)
乙醇的化学方程式2.5.4 其他多种卷积运算 (19)
2.6 本章小结 (23)
第三章轻量级去雾网络LMDehaze-Net的实现 (24)
3.1 概述 (24)
3.2 经典大气光散射物理模型的改进 (24)
3.3 LMDehaze-Net的网络结构 (25)
3.3.1 多尺度的卷积 (25)
3.3.2 混色通道的多尺度卷积 (27)
3.3.3 单色通道的多尺度卷积 (27)
3.3.4 大气光散射改进模型的结合 (29)
3.3.5 LMDehaze-Net网络结构参数 (29)
3.4 LMDehaze-Net去雾模型与传统去雾算法的关联分析 (30)
3.4.1 传统方法的相关去雾特征 (30)
3.4.2 LMDehaze-Net与传统去雾算法的关联分析 (31)
3.5 本章小结 (32)
第四章LMDehaze-Net网络模型相关实验 (33)家常炖猪肉
4.1 概述 (33)
4.1.1 实验数据集 (33)
4.1.2 实验评价指标 (34)
4.2 与其他算法在图像去雾方面的对比实验 (35)
4.2.1 实验结果分析 (37)
4.3 大气光散射改进模型融入前后的对比实验 (38)
4.3.1 实验结果分析 (38)
4.4 LMDehaze-Net在不同损失函数情况下的对比实验 (38)
4.4.1 损失函数的介绍 (38)
4.4.2 实验结果分析 (39)
4.5 本章小结 (39)
第五章模型部署Jetson™ TX1及相关基础应用的实现 (41)
5.1 概述 (41)
5.2 可移动端NVIDIA® Jetson™ TX1 (41)
5.3 基于Jets on™ TX1模型部署及应用总体流程 (43)拍手游戏
5.4 基于Jetson™ TX1的深度学习环境搭建 (44)
5.5 基于Jetson™ TX1的单张图像处理 (44)
5.6 基于Jetson™ TX1的LMDehaze-Net模型相关基础应用的实现 (45)
5.6.1 图片去雾模式 (46)
5.6.2 录像去雾模式 (47)
