结合支持向量机的卷积神经网络分类
结合支持向量机的卷积神经网络分类(论文11000字,外文翻译)
摘要：卷积神经网络（CNN）类似于“普通”神经网络，它是由具有“可学习”参数的神经元组成隐藏层组成。这些神经元操作过程是接收输入，执行点积，之后以非线性跟随它。整个网络表示原始图像与其识别分数之间的映射。一般的卷积神经网络通常是用Softmax函数在该网络的最后一层作为分类器使用。然而已经进行了研究[2,3,1]来重新定义这个规范。引用线性支持向量机（SVM）作为最后一层在神经网络结构中的应用。这个项目实验是另一个新的创新，受到[11]的启发和引导。实验数据表明，CNN-SVM模型能够使用MNIST数据集[10]达到“99.21％”的训练 测试精度。与此同时，CNN-Softmax使用相同的MNIST数据集能够达到“99.26％”的测试精度。两个模型也最新公布的Fashion-MNIST数据集[13]上进行了同样的测试，Fashion-MNIST数据集是比MNIST更难的图像分类数据集[15]，同样条件下更加难以识别。经过实验数据可以得出，CNN-SVM的测试精度达到了91.77％，而CNN-Softmax的测试精度达到了92.19％。Fashion-MNIST和CNN-SVM模型在相同数据集上基本可以达到相似的测试效果相差不到1%。
关键词：人工智能；卷积神经网络；Softmax函数；机器学习；支持向量机 [资料来源：Doc163.com]
Convolution neural network classification combined with support vector machine
Abstract：The Convolutional Neural Network (CNN) is similar to a "normal" neural network, which consists of a hidden layer of neurons with "learnable" parameters. These neuron operations are receiving input, performing a dot product, and then following it with nonlinearity. The entire network represents the mapping between the original image and its recognition score. A general convolutional neural network is usually used as a classifier in the last layer of the network using the Softmax function. However, research [2, 3, 1] has been carried out to redefine this specification. The linear support vector machine (SVM) is used as the last layer in the neural network structure. This project experiment is another new innovation, inspired and guided by [11]. Experimental data shows that the CNN-SVM model can achieve the "99.21%" training test accuracy using the MNIST data set [10]. At the same time, CNN-Softmax uses the same MNIST data set to achieve a "99.26%" test accuracy. The same test was performed on the two models of the recently published Fashion-MNIST dataset [13]. The Fashion-MNIST dataset is a more difficult image classification dataset than MNIST [15], which is more difficult to identify under the same conditions. The experimental data shows that the test accuracy of CNN-SVM has reached 91.77%, while the test accuracy of CNN-Softmax has reached 92.19%. The Fashion-MNIST and CNN-SVM models can achieve similar test results of less than 1% on the same data set. [来源：http://Doc163.com]
Key words：Artificial intelligence; Artificial neural network; Image classification; Machine learning; Support vector machine
[资料来源：www.doc163.com]
目 录
一、引言 5
（一）国内外研究现状 5
（二）本文研究内容 6
二、基础知识 6
（一）卷积神经网络基础知识 6
（二）支持向量机基础知识 13
三、结合支持向量机的卷积神经网络分类 14
（一）介绍 14
（二）方法 15
1.机器智能库 15
2.实验数据集 15
3.支持向量机（SVM） 16
4.卷积神经网络分类模型 16
（三）实验数据分析 18
1. CNN-SOFTMAX在MNIST数据集上的的训练及测试精度分析 19
2. CNN-SOFTMAX在Fashion-MNIST数据集上的的训练及测试精度分析 20
[来源：http://Doc163.com]
3. CNN-SVM在MNIST数据集上的的训练及测试精度分析 21
4. CNN-SVM在Fashion-MNIST数据集上的的训练及测试精度分析 22
四、实验对比分析和总结 23
（一）实验结果对比分析 23
1.在MNIST数据集上训练结果分析对比 23
2.在Fashion-MNIST数据集上训练结果分析对比 24
（二）实验总结 24
参考文献 26
致谢 28 [资料来源：http://Doc163.com]