基于深度学习的图像分类方法研究

发布时间：2018-01-02 03:17

本文关键词：基于深度学习的图像分类方法研究　出处：《中国矿业大学》2016年硕士论文　论文类型：学位论文

【摘要】：图像作为人类感知事物的视觉基础,是人们从外界获得信息的重要依据,所以让机器自动完成图像识别、分类具有重要意义。图像分类最重要的部分是特征提取,研究高效的特征提取算法在图像领域至关重要。深度学习(Deep Learning, DL)是多层的网络结构,它通过建立、模拟人脑的分层结构,对外部输入的声音、图像、文本等数据进行从低级到高级的特征提取,所以深度学习在图像分类领域具有广阔的应用空间。而深度学习本身存在训练时间过长、过拟合等问题,本文以提高深度模型分类精确度、缩短训练时间和防止模型过拟合三个问题为出发点,主要研究工作如下：首先,本文研究了极速学习机(Extreme Learning Machine,ELM)作为卷积神经网络(Convolutional Neural Network,CNN)分类器的可行性与意义,进而提出了混合深度模型CNN-ELM(Convolutional Neural Network-Extreme Learning Machine)。先用原始的CNN训练网络,然后用ELM替换CNN的输出层完成最后的分类,混合模型结合了CNN有效提取图像特征和ELM快速高效的特点,使得两种方法能够协同工作,实验表明CNN-ELM提高了CNN的分类精确度。其次,针对深度学习方法训练时间过长的问题,研究了随机参数网络结构的可行性与意义。核极速学习机是在ELM的基础上引入了核函数,具有更好的分类效果,从而提出了基于核极速学习机的随机参数深度模型：卷积极速学习机(Convolutional Extreme Learning Machine with Kernel,CKELM)。在模型CKELM中,把随机权值的卷积层和降采样层作为隐含层,来提取输入图像的显著特征。实验表明该算法既保证了分类精确度又大大缩短了深度算法的训练时间。最后,本文研究了基于DropConnect的深度自动编码器算法的应用意义和可行性。DropConnect作为一种新型的正则化方法,在处理过拟合等问题上表现突出,所以文章提出了一种基于DropConnect的深度自动编码器模型DDAE(DropConnect Deep AutoEncoder)。实验表明将DropConnect思想引入自动编码器中有效的提高了算法的性能。
[Abstract]:As the visual basis of human perception, image is an important basis for people to obtain information from the outside world, so the machine can automatically complete image recognition. Classification is of great significance. The most important part of image classification is feature extraction. It is very important to study efficient feature extraction algorithm in the field of image. DL) is a multi-layer network structure, it builds, simulates the human brain's hierarchical structure, carries on the low-level to the high-level feature extraction to the external input sound, the image, the text and so on data. So depth learning has a wide application space in the field of image classification, and depth learning itself has the problems of too long training time and over-fitting, so this paper improves the accuracy of depth model classification. Shortening the training time and preventing the model from overfitting is the starting point. The main research work is as follows: first. This paper studies extreme Learning Machine. The feasibility and significance of ELM as Convolutional Neural Network classifier. Furthermore, the mixed depth model CNN-ELM (. Convolutional Neural Network-Extreme Learning Machine. First use the original CNN to train the network. Then the final classification is completed by replacing the output layer of CNN with ELM. The hybrid model combines the features of CNN extraction and the fast and efficient feature of ELM, which makes the two methods work together. Experiments show that CNN-ELM improves the classification accuracy of CNN. Secondly, the training time of deep learning method is too long. The feasibility and significance of random parameter network structure are studied. The kernel function is introduced into the kernel pole learning machine based on ELM, which has better classification effect. Thus a random parameter depth model based on kernel pole learning machine is proposed: convolution extreme speed learning machine (. Convolutional Extreme Learning Machine with Kernel. In model CKELM, the convolution layer and downsampling layer of random weights are used as hidden layers. Experiments show that the algorithm not only ensures the classification accuracy but also greatly reduces the training time of the depth algorithm. This paper studies the application significance and feasibility of depth automatic encoder algorithm based on DropConnect. DropConnect is a new regularization method. In dealing with problems such as fitting outstanding performance. In this paper, a depth automatic encoder model DDAE(DropConnect Deep AutoEncoder based on DropConnect is proposed. The experimental results show that the performance of the algorithm is improved effectively by introducing the DropConnect idea into the automatic encoder.
【学位授予单位】：中国矿业大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【参考文献】