【计算机科学】【2015.05】基于卷积神经网络的鲁棒分类

本文为美国密苏里大学（作者：MUHIND SALIM HMOUD ALRADAD）的硕士论文，共64页。

对于图像分类，我采用了基于卷积神经网络（CNN）的方法，取得了目前最先进的性能。人脸图像分析需要有效的特征提取和分类器系统。本研究考虑深度学习演算法，并探讨其在分类工作中的应用。我展示了如何在CNN架构中选择平均和最大池化算法。该方法通过使用将平均和最大池化结合在一起的齐次网络来改进结果。对于基于人脸特征的性别分类，CNN可以有效地同时提取相关特征并进行分类。在两个无约束的数据集上获得了最先进的性能，即：野外人脸标记（LFW）和人群图像数据集。我们使用CAS PEAL-RL数据集在实验室内收集所有图像的约束条件下对我们的系统进行测试。对于年龄估计，我使用人群图像数据集取得了良好的性能，其中图像中的人群根据年龄被分为七组。

For image classification, I achieved thecurrent state-of-the-art performance by using methods based on ConvolutionalNeural Networks (CNNs). Face image analysis requires both effective featureextraction and classifier systems. This research considers the deep learningalgorithm and addresses its working for classification tasks. I shows how tochoose between averages and max pooling in CNN architecture. This methodimproved results by using homogeneous networks that combined average and maxpooling together. For gender classification based on facial features, CNNsproved effective for simultaneously extracting relevant features andclassifying them. State-of-the-art performance was obtained on twounconstrained datasets: Labeled Faces in the Wild (LFW) and Images of Groups ofpeople dataset. CASPEAL-RL dataset was used to test our systems underconstrained conditions where all the images were collected inside the lab. Forage estimation, I achieved good performance using images of groups of peopledataset where the people in the images have been divided into seven groups accordingto their age.

1.     引言
   

2.     项目背景
   

3.     性别与年龄分类
   

4.     结论与展望
   

下载地址：

https://url92.ctfile.com/f/1850492-516325048-1f0498

（访问密码：3660）

更多精彩文章请关注公众号：