基于卷积神经网络的图像分类-应用统计专业论文.docxVIP

研究生优秀毕业论文 大连理工大学专业学位硕士学位论文摘 大连理工大学专业学位硕士学位论文 摘 要 深度学习(Deep learning，DL)是机器学习(Machine Learning，ML)的一个崭新 的领域，它的目的就是让机器学习更加接近其终极目标．人工智能。卷积神经网络是深 度学习算法中的一种，具有结构简单、适应性强、训练参数少而连接多等特点，近年来 被广泛应用在图像处理和模式识别等领域。Dropout是一种训练神经网络的思路，通过 随机地将隐藏层的部分神经元的输出值归零而阻碍特征间的共生作用，降低了神经网络 模型的测试误差，进而提高了神经网络模型泛化能力。支持向量机是一个分类算法，通 过结构风险最小化理论提高模型的泛化能力。Caffe是一个开源的深度学习框架，具有 执行速度快和模块清晰等特点。本文提出了一个卷积神经网络模型MNIST-Net，在模型· 的最后一层采用Hinge Loss取代传统的Soft．Max回归进行分类，在未使用Dropout情况 下将MNIST测试集的峰点准确率从99．05％提高到99．36％，平均峰点准确率从98．964％ 提高到99．278％；在使用Dropout的情况下将峰点准确率从99．14％提高到99．39％，平 均峰点准确率从99．024％提高到99．321％。 关键词：卷积神经网络；支持向量机；Dropout；Caffe 万方数据 基于卷积神经网络的图像分类Image 基于卷积神经网络的图像分类 Image Classification based OI1 Convolutional Neural Networks Abstract Deep learning iS a new area of Machine learning research．which has been introduced with the objective of moving machine learning closer to one of its original goals：Artificial Intelligence．Convolutional neural networks(CNNs)，which is one of deep learning algorithms that is widely used in image processing and pattern recognition with the characteristics of simplicity，strong adaptability and few parameters．Dropout is a novel neural network training method，which randomly omits haIf of the feature detectors on each training case to reduce test error and improve the generality of the networks．Support vector machine is one classification algorithm，which improves the generalization ability of model through structural risk minimization．Caflfe iS a deep learning frame．work with expression，speed，and modularity in mind．In this article，we introduced a new CNN model using hinge lOSS as the last layer instead of traditional soft．Max regression and improved top accuracy from 99．05％ to 99。36％and average top accuracy from 98．964％to 99．278％without dropout and top accuracy from 99．1 4％to 99．39％and top accuracy from 99．024％to 99．32 1％with dropout． Key Words：CNNs；SVM；Dropout；Caffe II 万方数据 大连理工大学专业学位硕士学位论文目 大连理工大学专业学位硕士学位论文 目 录 摘 要 ．I Abstract ．． ．． ．．．．． ．．．．．．．．．．．． ． ． ． ．．．．．．．．．． ．．．．．．