【作者】 马健；

【导师】 李坤；

【作者基本信息】 天津大学 ， 模式识别与智能系统， 2018， 硕士

【摘要】 随着工业的发展,人们的生活质量大大提高,但随之而来的问题是大量的气溶胶,如灰尘,硫酸,硝酸和其他颗粒-雾霾,对人们的生活产生一定的负面影响。目前,近年来所有对空气污染水平进行分类的方法都是基于一定的污染参数,例如PM_2.5,PM₁₀等。PM_2.5指在大气中空气动力学粒径小于或等于2.5微米的颗粒物质。由于弥散在整个空气中,它可以被人们吸进入肺部从而带来健康隐患。实时,高效,低成本的估计大气污染对于人们日常生活是非常有必要的。本文围绕着基于图像的污染等级预测,通过深度卷积神经网络以及有监督学习方法展开研究,其主要创新点如下。（1）本文提出了一种利用深度混合卷积神经网络从单个图像中进行空气污染估算的方法。拍摄的图像输入到主网络,这是一个非常深的网络,通过跳过连接解决了增加深度（梯度消失或爆炸）的副作用。其次,计算暗通道图并将其馈入二级网络以使用隐式表示来丰富特征。合成数据集和实际自然场景下数据集的实验结果表明,我们的方法在基于图像的空气污染水平分类中实现了优异的性能。此外,基于深度混合卷积神经网络提取的高维特征,并利用支持向量回归（SVR）学习该特征与PM_2.5的映射关系。给定室外拍摄图像,本文算法可以实现实时对PM_2.5进行有效的监控。（2）本文提出了一种基于PM_2.5的雾霾天气下的单目图像深度估计方法。几乎所有现有的单目深度估计算法都侧重于干净无污染的图像,对于恶劣天气下,例如雾,霾等的数据图像却不适用。为了估计捕获场景的深度,我们首先利用稀疏的先验核和非局部双边核计算传输图,然后利用PM_2.5的估计值通过大气散射模型估计深度。实验结果表明,该方法可以达到与商品测量装置相同的PM_2.5估计精度,并且可以估计出比激光在无雾条件下捕获的真实数据更可信的深度信息。更多还原

【Abstract】 With the development of industry,people’s quality of life is greatly improved,but the problem is that a large amount of aerosols,such as dust,sulfuric acid,nitric acid and other aerosols-haze,have a certain negative impact on people’s lives.At present,all methods for classifying air pollution levels in recent years are based on certain pollution parameters,such as PM_2.5,PM₁₀,etc.PM_2.5refers to particulate matter with aerodynamic particle size of less than or equal to 2.5 microns in the atmosphere.It also becomes a particulate matter that can enter the lungs.It has a small particle size and is rich in a large amount of toxic and harmful substances.Real-time,efficient,low-cost estimation of atmospheric pollution is very necessary for people’s daily lives.This paper focuses on image-based pollution level prediction,through deep convolutional neural networks and supervised learning methods.The main innovations are as follows.（1）This paper proposes a method for estimating air pollution from a single image using a deep hybrid convolutional neural network,for example,captured by a smart-phone.The captured image is input to the main network,which is a very deep network that resolves the side effects of increasing depth（gradient disappearance or explosion）by skipping connection.This can improve network performance by simply increasing the depth of the network.Dark channel maps are computed and fed into the secondary network to enrich the features using implicit representations.We use different PM_2.5values??to train the end-to-end network.Experimental results of synthetic data sets and actual captured data sets show that our method achieves excellent performance in the classification of air pollution levels for a single captured image.In addition,the high-level features are extracted based on the deep mixed convolutional neural network,and the mapping relationship between the features and PM_2.5is learned by using sup-port vector regression.Real-time estimation of PM_2.5can be achieved given a captured image.（2）This paper proposes a method for estimating the monocular depth in foul weather.Almost all existing monocular depth estimation algorithms focus on clean,pollution-free images,but not for data images in harsh weather such as fog,haze,etc.To estimate the depth of the captured scene,we first compute the transmission map us-ing sparse prior and non-local bilateral kernel,and then estimate the depth through the atmospheric scattering model with the estimated PM_2.5.Experimental results demon-strate that the proposed method achieves the same level of PM_2.5-estimation accuracy as commodity measuring device,and can estimate plausible depth information that is even better than the“ground-truth”captured by a laser in the no-haze condition.This could be very useful in many applications under polluted air conditions.更多还原