电子学报 ›› 2020, Vol. 48 ›› Issue (10): 1969-1975.DOI: 10.3969/j.issn.0372-2112.2020.10.014

• 学术论文 • 上一篇    下一篇

一种基于可信度估计单元的图像分类噪声抑制深度学习策略

邵航1, 黄海亮1, 郭雨晨2, 戴琼海2   

  1. 1. 浙江未来技术研究院(嘉兴), 浙江嘉兴 314000;
    2. 清华大学自动化系, 北京 100084
  • 收稿日期:2019-06-10 修回日期:2020-03-22 出版日期:2020-10-25
    • 作者简介:
    • 邵航 男,1986年9月出生,浙江江山人.2011年毕业于清华大学自动化系,获工学硕士学位.主要研究方向为人工智能与计算成像.E-mail:shaohang@tsinghua.edu.cn
      黄海亮 男,1995年6月出生,江苏如皋人.2019年毕业于杭州电子科技大学自动化学院,获硕士学位.主要研究方向为人工智能与计算成像.E-mail:huanghailiang@zfti.org
    • 基金资助:
    • 国家自然科学基金 (No.61827805)

Noise-Suppression Deep Learning Strategy via Reliability Estimation Gate for Image Classification

SHAO Hang1, HUANG Hai-liang1, GUO Yu-chen2, DAI Qiong-hai2   

  1. 1. Zhejiang Future Technology Institute(Jiaxing), Jiaxing, Zhejiang 314000, China;
    2. Department of Automation, Tsinghua University, Beijing 100084, China
  • Received:2019-06-10 Revised:2020-03-22 Online:2020-10-25 Published:2020-10-25
    • Supported by:
    • National Natural Science Foundation of China (No.61827805)

摘要: 近年来,深度学习越来越关注噪声抑制的研究.本文提出了一种噪声抑制深度学习策略,该策略通过构建噪声无感网络(Noise Unaware Network,NUN)和可信度估计单元(Reliability Estimation Gate,REG)来处理训练数据含有噪声的情况.通过对每个样本的可信度进行评估,调节其在训练时的权重,从而降低标签噪声对网络训练的影响.随着模型的迭代更新,标签可信数据的权重将会逐渐变大,而噪声数据的权重将会被抑制.本文通过在多个标注数据集上的实验验证了所提出的噪声抑制深度学习策略的有效性.

关键词: 深度学习, 图像分类, 噪声抑制

Abstract: In recent years, more and more attention has been paid to the study and research of noise suppression. This paper proposes a noise suppression deep learning strategy that deals with the presence of noise in training data by constructing a Noise Unaware Network (NUN) and Reliability Estimation Gate (REG). By evaluating the reliability of each sample and adjusting its weight during training, the influence of label noise on network training can be reduced. As the model is updated iteratively, the weight of clean data will gradually increase, while the weight of noise data will be suppressed. Experiments on multiple benchmark data sets demonstrate the effectiveness of the proposed deep learning strategy for noise suppression.

Key words: deep learning, image classification, noise-suppression

中图分类号: