【作者】 黄勇；

【导师】 郭剑辉；

【作者基本信息】 南京理工大学 ， 计算机技术（专业学位）， 2020， 硕士

【摘要】 随着计算机发展的普及,数据也随着5G网络的出现而呈现出的激增的态势,传统的浅层学习算法在大样本数据下表现出计算能力不足以及训练效果较差等缺陷,深度学习是由多层神经网络构成的网络结构,具有强大的数据拟合能力,对于大量复杂的数据有极强的学习能力,可以较好的解决浅层学习算法应对复杂数据学习能力不足的问题,近年来,将深度学习算法与浅层学习算法的结合研究热点。本文主要研究深度信念网络(DBN)与支持向量机(SVM)并将该技术应用于地形识别领域。主要研究内容与提出观点如下:(1)分析了图像识别的理论与特征,对当前图像特征的提取和图像分类方法进行了详细的介绍。以及支持向量机的原理、局部二值模式(LBP)方法原理,以及与深度信念网络结合。同时研究了当前分类算法的主要评价标准。(2)对深度信念网络(DBN)与支持向量机(SVM)的方法应用在地形识别领域并提出了改进。先将图像通过局部二值模式进行特征提取,并将提取出的特征作为输入数据传入到初始构建的DBN网络结构进行学习得到更高级的特征表达,将DBN最后一层的隐含层输出数据作为SVM的输入数据进行学习。实验数据采用OUTEX与VSPECT数据集,通过实验验证,结合深度信念网络和支持向量机较传统浅层学习算法的有更好的精确度。(3)对稀疏编码与深度信念网络算法结合并应用于地形识别领域并提出了改进。在图像识别算法上,稀疏编码(SC)可以去除整体图像中的一些冗余信息,简化运算流程提高算法性能。同时稀疏编码在处理一些非线性数据时,可以提高图像识别的精确度,将其与深度信念网络结合后应用在图像识别领域。实验数据采用OUTEX与VSPECT数据集,通过实验验证,结合稀疏编码的深度信念网络算法的有效性。(4)结合随机隐退的深度信念网络算法并应用于地形识别领域并提出了改进。随机隐退(Dropout)通常在大网络结构训练小数据集的情况下使用,当样本数据较小时深度信念网络训练得到的网络模型可能导致局部收敛或者过拟合的情况。提出针对小样本情况下深度信念网络与随机隐退相结合的算法,实验数据采用OUTEX与VSPECT数据集,通过实验验证,结合随机隐退的深度信念网络算法的有效性。更多还原

【Abstract】 With the popularity of computer,data with the appearance of the 5G network also shows a tendency of the surge,traditional shallow learning algorithm under large sample data show the computation ability and the training effect is some shortcomings,deep learning network structure,is composed of multilayer neural network has strong ability of data fitting,for a large number of complex data have strong learning ability,can better solve the shallow learning algorithm ability to cope with complex data to study the problem of insufficient,in recent years,the deep learning algorithm combined with the shallow learning algorithm research hot spot.In this paper,deep belief network(DBN)and support vector machine(SVM)are studied and applied to terrain recognition.The main research contents and opinions are as follows:(1)The theory and features of image recognition are analyzed,and the current methods of image feature extraction and image classification are introduced in detail.And the principle of support vector machine,the principle of local binary mode(LBP)method,and the combination with deep belief network.At the same time,the main evaluation criteria of the current classification algorithm are studied.(2)The methods of deep belief network(DBN)and support vector machine(SVM)are applied in the field of terrain recognition and improvements are proposed.First,the image features are extracted through local binary mode,and the extracted features are passed as input data into the DBN network structure constructed initially for learning to obtain more advanced feature expression.Then,the output data of the hidden layer in the last layer of DBN is learned as the input data of SVM.OUTEX and VSPECT data sets were used for experimental data,and it was verified that the combination of deep belief network and support vector machine had better accuracy than the traditional shallow learning algorithm.(3)The sparse coding and deep belief network algorithm are combined and applied in the field of terrain recognition and improvements are proposed.In the image recognition algorithm,sparse coding(SC)can remove some redundant information in the whole image,to simplify the operation process and improve the algorithm performance.At the same time,sparse coding can improve the accuracy of image recognition when processing some nonlinear data,and it can be applied in the field of image recognition after combining it with deep belief network.OUTEX and VSPECT data sets were used to verify the validity of the deep belief network algorithm combined with sparse coding.(4)Combined with the depth belief network algorithm of random retreat and applied to the field of terrain recognition and proposed improvements.Dropout is usually used when large network structures are trained with small data sets.When the sample data is small,the network model obtained by deep belief network training may lead to local convergence or overfitting.An algorithm combining deep belief network and dropout was proposed for small samples.OUTEX and VSPECT data sets were used for experimental data,and the effectiveness of the algorithm combined with random withdrawal was verified through experiments.更多还原