發(fā)布時(shí)間：2018-07-28 07:25

【摘要】：深度學(xué)習(xí)是人工智能領(lǐng)域發(fā)展的一個(gè)重要組成部分,深度學(xué)習(xí)在許多領(lǐng)域(如圖像識(shí)別、語(yǔ)音識(shí)別、自然語(yǔ)言處理)取得了突破,在傳統(tǒng)算法不易解決的應(yīng)用方面也取得了令人可喜的成就,包括自動(dòng)無(wú)人駕駛汽車(chē)、自動(dòng)模式識(shí)別、自動(dòng)同聲傳譯、商品圖片檢索、手寫(xiě)字符識(shí)別、車(chē)牌自動(dòng)識(shí)別等。近年來(lái),隨著研究開(kāi)發(fā)人員對(duì)于深度學(xué)習(xí)開(kāi)發(fā)過(guò)程要求的不斷提高,傳統(tǒng)的深度學(xué)習(xí)編程方法已經(jīng)不能滿(mǎn)足當(dāng)前的需要,傳統(tǒng)的深度學(xué)習(xí)編程方法會(huì)耗費(fèi)研究開(kāi)發(fā)人員數(shù)月甚至幾年的時(shí)間用來(lái)實(shí)現(xiàn)最基本的算法,與此同時(shí),一些世界頂尖的科研機(jī)構(gòu)開(kāi)始尋求快速、高效的深度學(xué)習(xí)開(kāi)發(fā)模式,因此就產(chǎn)生了包括本文研究的Caffe深度學(xué)習(xí)框架在內(nèi)的多種深度學(xué)習(xí)開(kāi)發(fā)框架。這些深度學(xué)習(xí)框架不僅為科研機(jī)構(gòu)、相關(guān)開(kāi)發(fā)人員提供了高效、快速的開(kāi)發(fā)模式,并且其中一些深度學(xué)習(xí)框架還提供了多個(gè)卷積神經(jīng)網(wǎng)絡(luò)模型以便開(kāi)發(fā)人員在較為先進(jìn)、完善的卷積神經(jīng)網(wǎng)絡(luò)模型上進(jìn)行研究、改進(jìn)。本文基于Caffe深度學(xué)習(xí)框架的卷積神經(jīng)網(wǎng)絡(luò)進(jìn)行了以下幾項(xiàng)工作和研究:首先,介紹了關(guān)于深度學(xué)習(xí)在圖像識(shí)別,語(yǔ)音識(shí)別,自然語(yǔ)言處理三個(gè)領(lǐng)域的研究現(xiàn)狀,以及幾個(gè)主流的深度學(xué)習(xí)框架,并且進(jìn)行了比較,由此引出Caffe深度學(xué)習(xí)框架。之后,從人工神經(jīng)網(wǎng)絡(luò)引出卷積神經(jīng)網(wǎng)絡(luò),對(duì)卷積神經(jīng)網(wǎng)絡(luò)的構(gòu)成要素和結(jié)構(gòu)進(jìn)行了詳細(xì)闡述,對(duì)Caffe深度學(xué)習(xí)框架的幾個(gè)特性做了介紹,詳細(xì)說(shuō)明了Caffe環(huán)境搭建的步驟。最后,本文使用Caffe深度學(xué)習(xí)框架進(jìn)行仿真實(shí)驗(yàn)。仿真包括三個(gè)部分:1、以CIFAR-10的神經(jīng)網(wǎng)絡(luò)為例,對(duì)Caffe框架給出的卷積神經(jīng)網(wǎng)絡(luò)示例的配置訓(xùn)練方法進(jìn)行了說(shuō)明;2、以自己構(gòu)建的小型數(shù)據(jù)集為例,介紹了使用自己創(chuàng)建的數(shù)據(jù)集和自己搭建的卷積神經(jīng)網(wǎng)絡(luò)的訓(xùn)練方法;3、本文對(duì)Caffe框架下的基于MNIST手寫(xiě)字符集的LeNet-5網(wǎng)絡(luò)的改進(jìn),本文對(duì)激活函數(shù)進(jìn)行了改進(jìn),用ReLU函數(shù)替代了原始的Sigmoid,并在Le Net-5中加入了一層激活函數(shù),通過(guò)對(duì)比,本文的方法使得網(wǎng)絡(luò)的收斂速度有所提高,并且提高了網(wǎng)絡(luò)訓(xùn)練的正確率。
[Abstract]:Depth learning is an important part of the development of artificial intelligence. Deep learning has made a breakthrough in many fields (such as image recognition, speech recognition, natural language processing). Gratifying achievements have been made in the application of traditional algorithms, including autonomous driverless vehicles, automatic pattern recognition, automatic simultaneous interpretation, commodity image retrieval, handwritten character recognition, license plate automatic recognition and so on. In recent years, with the increasing demands of researchers and developers on the development process of in-depth learning, the traditional in-depth learning programming method can no longer meet the current needs. Traditional deep learning programming methods can take months or even years of research and development to implement the most basic algorithms, while some of the world's top scientific institutions are looking for fast, efficient models of deep learning and development. Therefore, many kinds of deep learning development frameworks, including the Caffe depth learning framework studied in this paper, have emerged. These deep learning frameworks not only provide efficient and rapid development models for scientific research institutions and related developers, but also provide multiple convolutional neural network models for developers to be more advanced. Perfect convolution neural network model is studied and improved. In this paper, the convolution neural network based on Caffe depth learning framework is studied as follows: firstly, the research status of depth learning in image recognition, speech recognition and natural language processing is introduced. And several mainstream deep learning frameworks are compared, which leads to the Caffe deep learning framework. After that, the convolution neural network is introduced from the artificial neural network, the composing elements and structure of the convolutional neural network are described in detail, several characteristics of the Caffe deep learning framework are introduced, and the steps of setting up the Caffe environment are explained in detail. Finally, this paper uses the Caffe depth learning framework to carry on the simulation experiment. The simulation includes three parts: 1. Taking the neural network of CIFAR-10 as an example, the configuration training method of convolutional neural network example given by Caffe framework is explained, and the small data set constructed by oneself is taken as an example. This paper introduces the training method of using the data set created by oneself and the convolution neural network built by oneself. In this paper, we improve the LeNet-5 network based on MNIST handwritten character set under the Caffe framework, and improve the activation function. The ReLU function is used to replace the original Sigmoid, and a layer of activation function is added to Le Net-5. By comparison, the convergence speed of the network is improved and the correct rate of network training is improved.
【學(xué)位授予單位】：河北師范大學(xué)
【學(xué)位級(jí)別】：碩士
【學(xué)位授予年份】：2016
【分類(lèi)號(hào)】：TP18

【參考文獻(xiàn)】

相關(guān)期刊論文 前1條

1 余凱;賈磊;陳雨強(qiáng);徐偉;;深度學(xué)習(xí)的昨天、今天和明天[J];計(jì)算機(jī)研究與發(fā)展;2013年09期

相關(guān)碩士學(xué)位論文 前2條

1 敖道敢;無(wú)監(jiān)督特征學(xué)習(xí)結(jié)合神經(jīng)網(wǎng)絡(luò)應(yīng)用于圖像識(shí)別[D];華南理工大學(xué);2014年

2 許可;卷積神經(jīng)網(wǎng)絡(luò)在圖像識(shí)別上的應(yīng)用的研究[D];浙江大學(xué);2012年

，

本文編號(hào)：2149341