本文選題：卷積神經(jīng)網(wǎng)絡 + 激活函數(shù)��； 參考：《北京交通大學》2017年碩士論文

【摘要】：卷積神經(jīng)網(wǎng)絡屬于深度學習領域研究的范圍,是一種高效的識別方法,卷積神經(jīng)網(wǎng)絡具有三個特點分別為參數(shù)共享,局部感知和子采樣操作,這三個特點使得訓練參數(shù)減少,訓練速度加快,在訓練過程中具有良好表現(xiàn),目前卷積神經(jīng)網(wǎng)絡已經(jīng)廣泛的并且良好的應用在生活各個方面,特別是在圖像分類任務,語音識別,文本識別,路標識別等方面。但其發(fā)展過程中還存在一些問題。本文將對卷積神經(jīng)網(wǎng)絡在圖像分類領域進行研究,目的是希望提高圖像分類的精準率,降低錯誤率。激活函數(shù)通過非線性函數(shù)把激活的神經(jīng)元的特征保留并映射出來,因此對于網(wǎng)絡性能有很大的影響,但是目前激活函數(shù)的選擇是一個問題,不同的激活函數(shù)具有不同的優(yōu)缺點,需要耗費大量的時間與精力來確定最優(yōu)的激活函數(shù)。本文主要針對激活函數(shù)選擇困難的問題,提出基于Relu-Softplus激活函數(shù)的卷積神經(jīng)網(wǎng)絡,并在手寫數(shù)字字體MNIST數(shù)據(jù)集上進行實驗,加以驗證其性能,并且同其他不同的激活函數(shù)進行比對,分析其圖像分類的錯誤率,以及收斂速度的快慢,最終達到優(yōu)化卷積神經(jīng)網(wǎng)絡的性能和解決確定最優(yōu)激活函數(shù)困難等問題的目的。卷積神經(jīng)網(wǎng)絡中的學習方式常見的有兩種,有監(jiān)督學習方法和無監(jiān)督學習方法,有監(jiān)督學習即從已標記的訓練樣本中學習到映射函數(shù),但是需要大量的訓練樣本,并且易出現(xiàn)過擬合等問題。而無監(jiān)督學習不要求訓練樣本帶有標簽,希望學習到更過抽象隱藏的特征結構,但具有訓練時間長,訓練過程繁瑣等缺點。本文主要針對此問題,提出基于K-means算法的卷積神經(jīng)網(wǎng)絡,并在CIFAR-10數(shù)據(jù)集上進行實驗,加以驗證其性能,并分析比較不同的網(wǎng)絡框架對圖像分類精準率的影響。最后本論文將卷積神經(jīng)網(wǎng)絡應用在路標識別系統(tǒng)上,并且設計了一個路標識別系統(tǒng),從系統(tǒng)的需求分析,概要設計,詳細設計以實現(xiàn)等方面進行了闡述。并將本文提出的基于K-means算法的卷積神經(jīng)網(wǎng)絡應用在路標識別系統(tǒng)中,最后在德國交通標志識別GTRSB數(shù)據(jù)集上進行訓練測試,并同其他知名的算法進行比較,加以驗證了基于K-means算法的卷積神經(jīng)網(wǎng)絡在路標識別系統(tǒng)的應用中對于路標分類的準確性,可靠性以及時效性方面確實有一定的提升。
[Abstract]:Convolutional neural network is an efficient recognition method, which belongs to the field of deep learning. It has three characteristics: parameter sharing, local sensing and sub-sampling operation, which make the training parameters reduced. At present, convolution neural network has been widely used in all aspects of life, especially in image classification task, speech recognition, text recognition, road sign recognition and so on. However, there are still some problems in its development. In this paper convolution neural networks are studied in the field of image classification in order to improve the accuracy of image classification and reduce the error rate. The activation function preserves and maps the characteristics of the activated neuron through the nonlinear function, so it has a great influence on the network performance. But at present, the choice of the activation function is a problem, and different activation functions have different advantages and disadvantages. It takes a lot of time and effort to determine the optimal activation function. Aiming at the difficulty of selecting activation function, a convolutional neural network based on Relu-Softplus activation function is proposed in this paper. Experiments are carried out on the MNIST dataset of handwritten digital font to verify its performance. Compared with other activation functions, the error rate of image classification and the speed of convergence are analyzed. Finally, the performance of convolution neural network is optimized and the problem of determining the optimal activation function is solved. There are two common learning methods in convolutional neural networks: supervised learning and unsupervised learning. Supervised learning is learning mapping functions from marked training samples, but a large number of training samples are required. And easy to have problems such as fitting. But the unsupervised learning does not require the training samples to be labeled, hoping to learn more abstract and hidden feature structures, but it has the disadvantages of long training time and tedious training process. In order to solve this problem, a convolutional neural network based on K-means algorithm is proposed in this paper. Experiments are carried out on the CIFAR-10 dataset to verify its performance, and the effects of different network frameworks on the accuracy rate of image classification are analyzed and compared. Finally, this paper applies the convolution neural network to the signpost recognition system, and designs a signpost recognition system, which is described from the aspects of system requirement analysis, summary design, detailed design and so on. The convolutional neural network based on K-means algorithm is applied to the road sign recognition system. Finally, the training test is carried out on the GTRSB data set of traffic sign recognition in Germany, and compared with other well-known algorithms. It is verified that convolution neural network based on K-means algorithm can improve the accuracy, reliability and timeliness of road sign classification in the application of road sign recognition system.
【學位授予單位】：北京交通大學
【學位級別】：碩士
【學位授予年份】：2017
【分類號】：TP391.41;TP183

【參考文獻】

相關期刊論文 前6條

1 黃毅;段修生;孫世宇;郎巍;;基于改進sigmoid激活函數(shù)的深度神經(jīng)網(wǎng)絡訓練算法研究[J];計算機測量與控制;2017年02期

2 趙玲玲;楊輝華;劉振丙;潘細朋;;基于深度卷積神經(jīng)網(wǎng)絡的乳腺細胞圖像分類研究[J];中小企業(yè)管理與科技(下旬刊);2016年06期

3 孫艷豐;楊新東;胡永利;王萍;;基于Softplus激活函數(shù)和改進Fisher判別的ELM算法[J];北京工業(yè)大學學報;2015年09期

4 呂剛;郝平;盛建榮;;一種改進的深度神經(jīng)網(wǎng)絡在小圖像分類中的應用研究[J];計算機應用與軟件;2014年04期

5 趙雷;張延榮;;基于概率神經(jīng)網(wǎng)絡和K-means算法的納稅評估[J];河北工程大學學報(社會科學版);2011年01期

6 吳佑壽,趙明生;激活函數(shù)可調的神經(jīng)元模型及其有監(jiān)督學習與應用[J];中國科學E輯:技術科學;2001年03期

相關碩士學位論文 前8條

1 姜含露;基于卷積神經(jīng)網(wǎng)的高光譜數(shù)據(jù)特征提取及分類技術研究[D];哈爾濱工業(yè)大學;2016年

2 產(chǎn)文濤;基于卷積神經(jīng)網(wǎng)絡的人臉表情和性別識別[D];安徽大學;2016年

3 何云超;聚類算法和卷積神經(jīng)網(wǎng)絡在文本情感分析中的應用研究[D];云南大學;2016年

4 張興革;基于卷積神經(jīng)網(wǎng)絡模型下的語音處理方法研究[D];東北林業(yè)大學;2016年

5 楊楠;基于Caffe深度學習框架的卷積神經(jīng)網(wǎng)絡研究[D];河北師范大學;2016年

6 吳正文;卷積神經(jīng)網(wǎng)絡在圖像分類中的應用研究[D];電子科技大學;2015年

7 岳永鵬;深度無監(jiān)督學習算法研究[D];西南石油大學;2015年

8 張凱歌;基于K-means和神經(jīng)網(wǎng)絡算法的圖像文字提取與識別[D];云南大學;2013年

，

本文編號：1775377