计算机与现代化

• 图像处理 • 上一篇    下一篇

强化深度特征融合的行人搜索系统

  

  1. (1.福州大学平板显示技术国家地方联合工程实验室,福建福州350116;
    2.福州大学物理与信息工程学院,福建福州350116)
  • 收稿日期:2019-01-18 出版日期:2019-08-15 发布日期:2019-08-16
  • 作者简介:梅文欣(1995-),女,福建宁德人,硕士研究生,研究方向:数字图像处理,机器学习,深度学习,E-mail: 17759371365@163.com; 林志贤(1975-),男,教授,博士,研究方向:信息显示技术,平板显示器件驱动,图像处理技术,E-mail: lzx2005000@163.com。
  • 基金资助:
    国家重点研发计划项目(2016YFB0401503); 广东省科技重大专项(2016B090906001); 福建省科技重大专项(2014HZ0003-1); 福建省资助省属高校专项课题(JK2014002)

Person Search System by Enhanced Deep Feature Fusion

  1. (1. National Joint Engineering Laboratory of Flat Panel Display Technology, Fuzhou University, Fuzhou 350116, China;
      2. College of Physics and Information Engineering, Fuzhou University, Fuzhou 350116, China)
  • Received:2019-01-18 Online:2019-08-15 Published:2019-08-16

摘要: 针对行人图像的深度特征缺乏对局部细节的描述,及不完全具备对尺度、旋转、平移及光照变化等各种因素的不变性而导致行人搜索准确率低的问题,本文提出一种具有强化深度特征融合的行人搜索系统。该系统将行人候选网络和行人识别网络两部分整合优化成统一框架。其中,行人候选网络实现行人框的获取及标定,而行人识别网络在获取深度学习特征的基础上融入具有几何不变性的传统特征,建立一个强化深度特征融合网络模型。实验结果表明,本文采用强化深度特征融合的网络模型,在SSM数据集上检测并框出图片中的行人,其Top-1识别正确率高达80.7%,比单纯采用深度特征模型更具优越性。

关键词: 深度特征, 行人搜索, 特征融合, 行人框, 几何不变性

Abstract: The deep feature of pedestrian image lacks the description of local details, and it does not have the invariance of scale, rotation, translation and illumination changes fully, which leads to the low accuracy of person search. A pedestrian search system with enhanced depth feature fusion is proposed. The system integrates the pedestrian candidate network and the pedestrian identification network into a unified framework. Among them, the pedestrian candidate network realizes the acquisition and calibration of the pedestrian boxes, while the pedestrian recognition network integrates the traditional features with geometric invariance on the basis of acquiring the deep learning characteristics, which establishes a network model with enhanced deep feature fusion. The experimental results show that the network model with enhanced depth feature fusion detects and frames pedestrians in images on SSM dataset, and has a top rate of 80.7%, which is superior to the deep feature model.

Key words: deep feature, pedestrian search, feature fusion, pedestrian boxes, geometric invariance

中图分类号: