基于多尺度残差注意网络的轻量级行人属性识别算法
作者:
作者单位:

贵州大学大数据与信息工程学院

作者简介:

通讯作者:

中图分类号:

TP391.41

基金项目:

贵州省科学技术基金资助项目(黔科合基础-ZK[2021] 重点001)


Lightweight Pedestrian Attribute Recognition Algorithm based on Multiscale Residual Attention Network
Author:
Affiliation:

College of Big Data and Information Engineering,Guizhou University

Fund Project:

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    近年来,行人属性识别得益于深度学习的蓬勃发展也得到了广泛的研究。但是由于属性复杂且多样化、图像质量差、视角遮挡等困扰,难以捕获图像中的细粒度属性特征,具有很大的挑战性。基于深度学习,我们提出了多尺度残差注意网络(MRAN)用于行人属性识别,以 Resnet50 为主体架构,使用轻量级的金字塔卷积提供不同内核大小的并行卷积完成多尺度信息的提取,嵌入注意力模块以关注属性存在的关键区域并挖掘属性内部联系;其次,使用特征金字塔融合策略,更充分提取和融合多尺度特征。网络结合了多尺度学习、注意力机制和残差学习的思想,使网络提取出更丰富、更细腻的特征。最后,在 PETA 和 PA100K 两个数据集上进行了实验研究,结果表明,该方法优于现有的研究方法。通过消融研究验证了整个网络体系结构的三个组成部分的有效性和先进性,且所提网络具有高准确性和低复杂度的双向优化。

    Abstract:

    Recently, pedestrian attribute recognition has been extensively studied that has benefited from the vigorous development of deep learning. However, it is difficult to capture the fine-grained attributes in the image due to complex and diversified attributes, poor image quality, and viewing angle occlusion, which is very challenging. Based on deep learning, we propose a Multi-scale Residual Attention Network (MRAN) for pedestrian attribute recognition with Resnet50 as the main architecture, using lightweight pyramid convolution to provide parallel convolution with different kernel sizes to complete multi-scale information extraction. The attention module is embedded to focus on the key areas where the attributes exist and explore the internal relations of the attributes. Secondly, the feature pyramid aggregation strategy is used to more fully extract and fuse multi-scale features. The network combines the ideas of multi-scale learning, attention mechanism and residual learning to enable the network to extract richer and more delicate features. Finally, an experimental study was carried out on two datasets of PETA and PA100K, and the results showed that proposed method is superior to the existing research methods. Through ablation research, the effectiveness and advancement of the three components of the entire network architecture are verified, and the proposed network has bidirectional optimization with high accuracy and low complexity.

    参考文献
    相似文献
    引证文献
引用本文
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2021-03-10
  • 最后修改日期:2021-06-25
  • 录用日期:2021-07-05
  • 在线发布日期: 2021-08-01
  • 出版日期: