基于卷积神经网络的图像识别

发布时间：2018-03-22 07:02

本文选题：深度学习　切入点：残差　出处：《吉林大学》2017年硕士论文　论文类型：学位论文

【摘要】：生物神经网络是由海量的神经元通过树突和轴突连接而成,其能够不断学习和进化。而人工神经网络就是基于生物学中神经网络的基本原理,以网络拓扑知识为基础,模拟人脑的神经系统对外界信息的处理机制的一种数学模型。人工神经网络的研究在上世纪末期经历了一段时间的快速发展后,由于各种条件的限制,陷入了低谷。深度学习就是以人工神经网络为基础,一种更为深层次并且更加复杂的网络模型。近年来,由于计算机技术的迅猛发展,尤其是硬件条件的提升,深层神经网络的训练时间大大减少,深度学习也迅速成为一个研究热点。在数字识别,语音识别,无人驾驶,图像识别等领域,深度学习技术都有着广泛的应用。以深度学习为代表的新一代人工智能技术,正在逐渐渗透到人们生活之中,并推动着社会的发展。而卷积神经网络作为深度学习技术的一个代表,近年来也取得了飞速的发展。卷积神经网络是一种特殊的深层神经网络,它包含特殊的卷积层和降采样层,卷积层的神经元和前一层采用局部连接和权值共享的方式进行连接,从而大大降低了需要训练的参数数量。降采样层可以大幅降低输入维度,降低网络复杂度,使网络具有更高的鲁棒性,同时能够有效的抑制过拟合问题。这些设计使得卷积神经网络在识别缩放,位移以及其他形式扭曲不变的二维图形的任务上有着无可比拟的优越性。2012年的Alex-net在imagenet的比赛上大放异彩,而2015年imagenet大赛上出现的残差神经网络,又是卷积神经网络的一种新的变形,它让卷积层不直接学习目标,而是转而学习一种残差。残差神经网络的出现,让卷积网络的研究又进入了一个新的阶段。本文基于传统的残差网络,提出了一种改进的多尺度残差网络,我们的网络结构使得卷积层能够从多个不同的尺度“观察”数据,获得更为丰富的输入特征。并且缩减了网络的深度,有效了抑制了梯度消失问题的发生,减小了训练难度。同时增加了可训练参数的数量,使得网络的学习能力变得强。我们通过调整缩放参数,网络深度,每组学习模块的数量,dropout的位置及取值,使得我们的网络结构的准确率比传统的残差神经网络有了明显的提升。最后,我们使用了相对多数投票的集成学习方法,将网络在CIFAR-10上的分类错误率降低到3.49%,相对于原始的残差神经网络,错误率下降了3%。
[Abstract]:Biological neural networks are composed of massive neurons connected by dendrites and axons, which can be continuously studied and evolved. Artificial neural networks are based on the basic principles of neural networks in biology and are based on the knowledge of network topology. A mathematical model that simulates the processing mechanism of the human brain's nervous system to external information. After a period of rapid development in the late last century, the study of artificial neural networks has been restricted by various conditions. Deep learning is based on artificial neural network, a deeper and more complex network model. In recent years, due to the rapid development of computer technology, especially the improvement of hardware, The training time of deep neural network is greatly reduced, and deep learning has become a research hotspot. In the fields of digital recognition, speech recognition, unmanned driving, image recognition and so on, Deep learning technology is widely used. The new generation of artificial intelligence technology, represented by deep learning, is gradually infiltrating into people's lives. As a representative of deep learning technology, convolution neural network has made rapid development in recent years. Convolutional neural network is a special kind of deep neural network. It contains special convolution layer and down-sampling layer. The neurons of the convolutional layer and the previous layer are connected by local connection and weight sharing, which greatly reduces the number of parameters that need to be trained. The complexity of the network is reduced, the network is more robust, and the over-fitting problem can be effectively suppressed. These designs make the convolutional neural network recognize scaling. The task of displacement and other forms of distorted two-dimensional graphics has unparalleled advantages. In 2012, Alex-net made a splash in the imagenet competition, and in the 2015 imagenet contest, the residual neural network emerged. It is a new kind of deformation of convolutional neural network, which makes the convolution layer not learn the target directly, but instead learn a kind of residuals. In this paper, an improved multi-scale residual network is proposed based on the traditional residuals network. Our network structure enables the convolution layer to "observe" data from many different scales. More abundant input features are obtained, and the depth of the network is reduced, the occurrence of the gradient vanishing problem is effectively restrained, and the difficulty of training is reduced. At the same time, the number of trainable parameters is increased. Make the learning ability of the network strong. We can adjust the scaling parameters, the depth of the network, the number of learning modules in each group and the location and value of dropout. This makes our network structure more accurate than the traditional residuals neural network. Finally, we use the ensemble learning method of relative majority voting. The classification error rate of the network on CIFAR-10 is reduced to 3.49. Compared with the original residual neural network, the error rate is reduced by 3%.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【相似文献】