首页> 外文学位 >Ensemble of Deep Neural Netwoks for Image Analysis
【24h】

Ensemble of Deep Neural Netwoks for Image Analysis

机译:深度神经网络集成的图像分析

获取原文
获取原文并翻译 | 示例

摘要

Deep learning has been able to achieve impressive results in recent years. But the availability of suitable amount of domain-specific data remains a challenge especially in image related tasks where dataset can vary in size from a few hundred images to millions of images. A sufficiently deep neural network with millions of parameters needs huge amount of data for training. On the other hand, training deep neural networks on a large dataset is computationally expensive. In this context, this thesis explores an ensemble learning based approach for image recognition. Multiple pre-trained Convolution Neural Networks (CNNs) are fine-tuned as base learners, and they are combined using a meta learner to improve the overall performance. This approach provides reasonable accuracy with comparatively less computational cost. As part of this study, a novel regression model is developed with 3 CNNs as base learners and a feed-forward neural network as meta-learner to predict the value of fine particulate matter (PM2.5) from image using a small image dataset. The experimental results demonstrate that the proposed method provides a more accurate PM 2.5 prediction compared to the individual CNNs and therefore it can be used for image-based PM2.5 estimation. A relatively similar approach is applied for a classification task with six CNNs as base learners using a large image dataset. In this case also, the ensemble-based approach outperforms the individual CNNs in terms of classification accuracy.
机译:近年来,深度学习已经取得了令人瞩目的成果。但是,提供适当数量的特定于域的数据仍然是一个挑战,尤其是在图像相关任务中,其中数据集的大小可能从几百个图像到数百万个图像不等。具有数百万个参数的足够深的神经网络需要大量的数据进行训练。另一方面,在大型数据集上训练深度神经网络的计算量很大。在这种背景下,本文探索了一种基于整体学习的图像识别方法。多个预训练的卷积神经网络(CNN)作为基础学习者进行了微调,并使用元学习器将它们组合在一起以提高整体性能。这种方法以较低的计算成本提供了合理的精度。作为这项研究的一部分,开发了一种新颖的回归模型,其中以3个CNN为基础学习器,并使用前馈神经网络作为元学习器,以使用小型图像数据集从图像中预测细颗粒物(PM2.5)的值。实验结果表明,与单个CNN相比,该方法可提供更准确的PM 2.5预测,因此可用于基于图像的PM2.5估计。相对相似的方法适用于使用大型图像数据集的六个CNN作为基础学习者的分类任务。同样在这种情况下,基于分类的方法在分类准确度方面也优于单个CNN。

著录项

  • 作者

    Rijal, Nabin Sharma.;

  • 作者单位

    Lamar University - Beaumont.;

  • 授予单位 Lamar University - Beaumont.;
  • 学科 Computer science.
  • 学位 M.C.Sc.
  • 年度 2018
  • 页码 56 p.
  • 总页数 56
  • 原文格式 PDF
  • 正文语种 eng
  • 中图分类
  • 关键词

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