基于深度学习的自然图像分类方法的研究

发布时间：2018-05-20 19:33

本文选题：深度学习 + 图像分类　；参考：《东华理工大学》2017年硕士论文

【摘要】：近几年以来,随着科学计算机网络及人工智能领域的发展,图形图像数据量逐渐增多,于是,如何从大量的自然图像中快速提取到视觉特征已经成了机器智能学习中的热点研究课题,进而对自然图像的分类必然成为获取自然图像信息的研究重点。卷积神经网络是深度学习在图像处理方面的一个重要应用。它相比于其它机器学习算法如SVM等,其优点是能够直接对图像像素进行卷积并提取特征,也能够利用海量的图像数据将网络参数训练充分,以达到更好的分类效果。本文对基于深度学习的自然图像分类方法展开研究,主要工作及创新点如下:1)基于tensorflow深度学习框架平台设计一个用于识别图像的浅层卷积神经网络,并分别用单GPU和多GPU训练加速来对比该网络性能,其中多GPU训练该网络的所用的时间比单GPU缩短了25分钟。该项工作的设计旨在建立一个较好的网络结构进行训练和评估,并为工作3中建立更加复杂的网络模型做铺垫。2)本文围绕卷积神经网络的网络结构和多参数分别进行了改进和优化。研究实验表明,对batch值、dropout、momentum动量值、数据集扩增等的优化,能够有效地提高深层卷积神经网络模型的识别率。因此,合理的增加网络层数,优化训练参数提高训练效率,以达到最佳的分类效果是图像分类应用研究中非常重要的目的。3)基于tensorflow深度学习框架平台,并用GPU训练加速来进行卷积神经网络的网络结构改进设计和参数优化。首先,设计一个具有9层结构的深层卷积神经网络。其次,用该网络结构分别对cifar-10和cifar-100等复杂图像数据库进行训练、测试和优化参数。结果表明,该网络结构相比之前研究者的网络模型(Conv-KN)对这两种复杂的图像库的分类准确率分别提高了9.26%和3.55%。tensorflow框架平台下的深层卷积神经网络的分类效果要明显好于其它平台,并且tensorflow框架平台下的训练时间上也得到了极大的提高。
[Abstract]:In recent years, with the development of scientific computer network and artificial intelligence, the amount of graphic and image data is increasing gradually. So how to quickly extract visual features from a large number of natural images has become a hot topic in machine intelligent learning, and then the classification of natural images will inevitably become the acquisition of natural image information. Research emphasis. Convolution neural network is an important application of deep learning in image processing. Compared with other machine learning algorithms, such as SVM, it has the advantage that it can convolution the image pixels directly and extract the features, and can also use massive image data to train the network parameters fully in order to achieve better classification results. This paper studies the classification method of natural image based on deep learning. The main work and innovation are as follows: 1) based on the tensorflow deep learning framework platform, a shallow convolution neural network for identifying images is designed, and the performance of the network is compared with single GPU and multiple GPU training respectively, in which the multi GPU is used to train the network. The time is 25 minutes shorter than the single GPU. The design of the work is designed to build a better network structure for training and evaluation, and to build a more complex network model for work 3. This paper improves and optimizes the network structure and multi parameters of the convolution neural network respectively. The research experiments show that the batch value, D The optimization of ropout, momentum momentum, and data set amplification can effectively improve the recognition rate of the deep convolution neural network model. Therefore, it is very important for the image classification application to increase the number of network layers, optimize the training parameters and improve the training efficiency, and to achieve the best classification effect in the study of the image classification application. Based on the depth of tensorflow, the.3 Study frame platform and accelerate the network structure improvement design and parameter optimization of convolution neural network with GPU training. First, a deep convolution neural network with 9 layers structure is designed. Secondly, the network structure is used to train, test and optimize the parameters of complex image databases such as cifar-10 and cifar-100, respectively. Compared with the previous researchers' network model (Conv-KN), the classification accuracy of the two complex image bases is improved by 9.26% and the deep convolution neural network under the 3.55%.tensorflow framework platform is better than the other platforms, and the training time under the tensorflow frame platform is also greatly improved. Improve.
【学位授予单位】：东华理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP18

【参考文献】