基于卷积神经网络的场景分类的研究与应用

发布时间：2018-10-26 11:35

【摘要】：场景分类是图像处理领域的重要研究方向之一。随着计算机技术和互联网的发展,大量的图像数据涌入到人们的生活和工作中,面对如此巨大的图像信息,传统的场景分类方法和技术表现出很多不足。近年来,卷积神经网络(Convolutional Neural Network,CNN)在图像处理领域取得了很多突破性进展,它是通过模拟人类大脑学习的过程,直接从图像像素中提取图像特征,并将特征提取与分类器结合到一个学习框架下,对相关对象进行分类识别。另外,卷积神经网络的局部连接、权值共享和降采样大大减少了网络的训练参数,简化了网络模型,进一步提高了网络的训练效率。本文针对场景图像的复杂多变性和传统场景分类方法泛化能力不强的问题,结合卷积神经网络方法进行场景分类。卷积神经网络分类性能的好坏主要决取于网络的层次结构,因此本文研究了影响卷积神经网络分类性能的因素,并以此为根据设计了一个卷积神经网络模型,应用于场景分类中。具体工作如下:1.针对应用于场景分类设计的卷积神经网络模型中如何选择层次结构问题,本文设计了一个浅层卷积神经网络模型,应用于Scene-15数据集和SUN-397数据集的场景图像分类任务中,以此研究不同大小和个数的卷积核、不同的激活函数和不同采样方法对卷积神经网络分类性能的影响。研究表明神经网络使用较小的卷积核以及较多的核数目、最大值采样和使用ReLU激活函数,可增加卷积神经网络的分类性能。2.为更好地适应实际场景图像的要求,本文根据以上研究对神经网络模型进行了改进,设计了一个8层的卷积神经网络。该网络的卷积层采用了较小的卷积核,并增加了卷积核的数量,这样可以提取到更多的图像特征,提高分类性能。同时,采样层采用了最大值采样方法以及ReLU激活函数。本文把改进后的卷积神经网络模型与AlexNet模型和VGGNet模型在Scene-15数据集和SUN-397数据集上进行了对比实验,实验结果证明了该模型在场景分类应用中具有良好的分类效果。本文主要是在MATLAB软件上利用MatConvNet工具箱进行卷积神经网络的结构设计和参数优化,分析了影响卷积神经网络分类性能的因素,并以此为根据设计了卷积神经网络模型,应用于场景分类中。大量实验表明本文网络模型在场景分类应用中具有良好的分类性能,并具有一定的泛化能力。
[Abstract]:Scene classification is one of the important research directions in the field of image processing. With the development of computer technology and the Internet, a large number of image data flow into people's lives and work. In the face of such huge image information, traditional scene classification methods and techniques show a lot of shortcomings. In recent years, convolutional neural network (Convolutional Neural Network,CNN) has made many breakthroughs in the field of image processing. It extracts image features directly from image pixels by simulating the learning process of human brain. The feature extraction and classifier are combined into a learning framework to classify and recognize the related objects. In addition, the local connection, weight sharing and down-sampling of convolutional neural networks greatly reduce the training parameters of the network, simplify the network model, and further improve the training efficiency of the network. Aiming at the complex variability of scene image and the weak generalization ability of traditional scene classification methods, this paper combines convolution neural network method to classify scene. The classification performance of convolutional neural networks is mainly determined by the hierarchical structure of the network. Therefore, the factors influencing the classification performance of convolutional neural networks are studied in this paper, based on which a convolution neural network model is designed. Applied to scene classification. The specific work is as follows: 1. Aiming at the problem of how to select hierarchical structure in the model of convolution neural network applied in scene classification design, a shallow convolution neural network model is designed in this paper, which is applied to the task of scene image classification in Scene-15 dataset and SUN-397 dataset. The effects of different size and number of convolution kernels, different activation functions and different sampling methods on the classification performance of convolution neural networks are studied. It is shown that the classification performance of convolutional neural networks can be improved by using smaller convolutional kernels and more kernel numbers, maximum sampling and ReLU activation function. 2. In order to better meet the requirements of the actual scene image, this paper improves the neural network model based on the above research, and designs an 8-layer convolution neural network. The convolution layer of the network uses smaller convolution cores and increases the number of convolution cores which can extract more image features and improve classification performance. At the same time, the maximum sampling method and the ReLU activation function are used in the sampling layer. In this paper, the improved convolution neural network model is compared with AlexNet model and VGGNet model on Scene-15 data set and SUN-397 data set. The experimental results show that the model has good classification effect in scene classification. In this paper, the structure design and parameter optimization of convolution neural network are carried out by using MatConvNet toolbox on MATLAB software. The factors influencing the classification performance of convolution neural network are analyzed, and the convolution neural network model is designed. Applied to scene classification. A large number of experiments show that the network model has good classification performance and generalization ability in scene classification.
【学位授予单位】：兰州理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【参考文献】