• 主题
高级搜索  >
历史搜索 清空历史
首页> 外文学位 >OneClass boosting and its application to classification problems.
【24h】

OneClass boosting and its application to classification problems.

机译：OneClass Boosting及其在分类问题中的应用。

获取原文
获取原文并翻译 | 示例
页面导航
  • 摘要
  • 著录项
  • 相似文献
  • 相关主题

摘要

We promote the concept of one-class learning and explore the role of background examples in learning processes. We propose to fit a stepwise additive model to minimize an adaptive quadratic loss. Experimental results on the NIST dataset show that our algorithm achieves comparable accuracy to AdaBoost, while it requires no individual training examples; robust to noise in the training data, and is computationally very efficient. Exponential loss is introduced to address the situation when there are small clusters in the object class. Randomization scheme is incorporated to handle high dimension of data. The resulted OneClassBoost converges much faster than AdaBoost, and achieves superior performance to AdaBoost in presence of small training sets. Behind the success of OneClassBoost, we believe that, simple statistics on background, rather than individual background examples, drive to shape the classification boundary.;Also presented in this work is a framework for rapid objection detection. Promising results have been obtained using hierarchical search and attentional cascades. Hierarchical search is an efficient way of organizing objects in order to accommodate a large number of possible hypotheses. Cascade is a technique which allows background regions of an image to be quickly discarded. Our approach is to unify the two designs. The detection process corresponds to a sequential testing which is coarse-to-fine in both the exploration of object poses and the representation of background. The average detection time is 10%--15% of that of a single cascade.;Both the one-class learning and the framework of test design is motivated by the task of face detection. Toward this end we have constructed a frontal face detection system which achieves an accuracy comparable to the state of the art on benchmark test sets, and capable of processing 7 frames per second.
机译：我们提倡一课式学习的概念，并探讨背景示例在学习过程中的作用。我们建议拟合逐步加法模型以最小化自适应二次损失。在NIST数据集上的实验结果表明，我们的算法可达到与AdaBoost相当的精度，而无需单独的训练示例。对训练数据中的噪声具有鲁棒性，并且计算效率很高。当对象类中存在小的簇时，引入指数损失来解决这种情况。随机方案被纳入处理高维数据。所得的OneClassBoost收敛速度比AdaBoost快得多，并且在存在小型训练集的情况下，其性能优于AdaBoost。在OneClassBoost成功的背后，我们相信，有关背景的简单统计而不是单个背景的示例会推动分类边界的形成。这项工作还介绍了一种快速异议检测的框架。使用分层搜索和注意级联已经获得了有希望的结果。分层搜索是组织对象以适应大量可能假设的有效方法。级联是一种允许快速丢弃图像背景区域的技术。我们的方法是统一这两种设计。检测过程与顺序测试相对应，该顺序测试在探究对象姿态和背景表示时从粗到细。平均检测时间是单个级联检测时间的10％-15％。一类学习和测试设计框架都是由人脸检测的任务驱动的。为此，我们构建了一个正面人脸检测系统，该系统的精度可与基准测试仪相媲美，并且能够每秒处理7帧。

著录项

  • 作者

    Xu, Qingqing.;

  • 作者单位

    The University of Chicago.;

  • 授予单位 The University of Chicago.;
  • 学科 Statistics.;Computer Science.
  • 学位 Ph.D.
  • 年度 2009
  • 页码 167 p.
  • 总页数 167
  • 原文格式 PDF
  • 正文语种 eng
  • 中图分类 统计学;自动化技术、计算机技术;
  • 关键词

  • 入库时间 2022-08-17 11:38:26

相似文献

  • 外文文献
  • 中文文献
  • 专利
获取原文

客服邮箱：kefu@zhangqiaokeyan.com

京公网安备：11010802029741号 ICP备案号：京ICP备15016152号-6 六维联合信息科技 (北京) 有限公司©版权所有

  • 客服微信

  • 服务号