计算机工程 ›› 2019, Vol. 45 ›› Issue (9): 242-247.doi: 10.19678/j.issn.1000-3428.0051855

• 图形图像处理 • 上一篇    下一篇

一种基于迁移学习的能见度检测方法

唐绍恩1, 李骞1, 胡磊2, 马强1, 顾大权1   

  1. 1. 国防科技大学 气象海洋学院, 南京 211101;
    2. 61741部队, 北京 100094
  • 收稿日期:2018-06-19 修回日期:2018-08-17 出版日期:2019-09-15 发布日期:2019-09-03
  • 作者简介:唐绍恩(1993-),男,硕士研究生,主研方向为图像检测;李骞(通信作者),副教授、博士;胡磊,工程师、硕士;马强,讲师、硕士;顾大权,教授
  • 基金项目:
    国家自然科学基金(41305138);中国博士后科学基金(2017M621700)。

A Visibility Detection Method Based on Transfer Learning

TANG Shaoen1, LI Qian1, HU Lei2, MA Qiang1, GU Daquan1   

  1. 1. College of Meteorology and Oceanography, National University of Defense Technology, Nanjing 211101, China;
    2. 61741 Troops, Beijing 100094, China
  • Received:2018-06-19 Revised:2018-08-17 Online:2019-09-15 Published:2019-09-03

摘要: 为学习可有效反映能见度的视觉特征,解决大规模训练数据集构建困难的问题,提出一种将深度卷积神经网络应用于能见度检测的方法。将样本图像划分为多个子区域,利用预训练的VGG-16网络对其进行编码。通过编码特征集训练支持向量回归模型,并根据支持向量误差计算各子区域的融合权重,按权重融合子区域能见度估计值。实验结果表明,该方法检测正确率超过90%,可满足实际应用的需求。

关键词: 能见度检测, 深度神经网络, 迁移学习, 支持向量回归, 权重融合

Abstract: In order to study the image features that can effectively reflect visibility,and solve the difficulties to structure a large-scale training data set,this paper proposes a method for applying deep convolution neural networks to visibility detection.The sample image is divided into several subdomains and encoded by using the pre-trained VGG-16 network.Support Vector Regression(SVR) models are trained with coded feature sets,and each subdomains' fusion weight is calculated according to the error analysis of the support vector,and then fuse the visibility estimates of subdomains by weight.Experimental results show that the detection accuracy of the proposed method exceeds 90%,which can meet the requirements of application.

Key words: visibility detection, deep neural networks, transfer learning, Support Vector Regression(SVR), weights fusion

中图分类号: