作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2024, Vol. 50 ›› Issue (6): 102-109. doi: 10.19678/j.issn.1000-3428.0068317

• 人工智能与模式识别 • 上一篇    下一篇

多尺度融合与双输出U-Net网络的行人重识别

胡庆   

  1. 东北石油大学三亚海洋油气研究院, 海南 三亚 572024
  • 收稿日期:2023-08-31 修回日期:2023-12-14 发布日期:2024-03-06
  • 通讯作者: 胡庆,E-mail:081980010077@nepu.edu.cn E-mail:081980010077@nepu.edu.cn
  • 基金资助:
    海南省贾承造、杨树锋院士工作站资金资助;海南省高新技术项目(ZDYF2023GXJS004);中国石油天然气集团有限公司科学研究与技术开发项目(2021DJ2504)。

Multi-Scale Fusion and Dual Output U-Net Network for Person Re-Identification

HU Qing   

  1. Sanya Institute of Offshore Oil and Gas, Northeast Petroleum University, Sanya 572024, Hainan, China
  • Received:2023-08-31 Revised:2023-12-14 Published:2024-03-06

摘要: 受行人姿态的多变性和行人被遮挡等因素的影响,行人重识别模型难以提取行人关键特征。为增强模型的特征表达能力,提出一种基于多尺度融合与双输出U-Net网络的行人重识别方法,旨在解决现有方法中难以提取行人关键特征、特征表达能力较低的问题。首先,提出多尺度融合的双输出U-Net网络,并对输出特征进行欧氏距离和散度距离约束;其次,设计联合损失函数,解决生成对抗网络在训练过程中不易收敛的问题,提高训练过程的收敛速度。在3个公共基准数据集上的仿真实验结果表明,相比经典特征提取网络,所提特征提取网络的平均精度均值(mAP)提升超过10%,所提行人重识别方法相比主流方法的mAP提高约2%,该方法能够增强模型的特征表达能力,提高行人重识别的准确率。

关键词: 行人重识别, 生成对抗网络, 特征提取, 多尺度融合, 联合约束

Abstract: Due to variable pedestrian posture, the occlusion of pedestrians, and other adverse factors, Person Re-Identification(Re-ID) models often struggle to extract the key features of pedestrians. To enhance the feature expression ability of the model, this study proposes a Re-ID method based on multi-scale fusion and a dual output U-Net network. The aim is to address the challenges of extracting key pedestrian features and improving feature expression, which are limitations of existing methods. First, a multi-scale fusion dual output U-Net network is proposed, with the output features constrained by both Euclidean and divergence distances. Second, a joint loss function is introduced to address the challenge of the Generative Adversarial Network(GAN) not converging easily during training, thereby improving the convergence speed of the training process. Numerous simulation experiments conducted on three public reference datasets demonstrate that the proposed feature extraction network outperforms classical feature extraction networks, with an improvement in Mean Average Precision (mAP) of over 10%. In addition, the mAP of the proposed Re-ID method improves by approximately 2% compared to that of the mainstream method. The proposed method can enhance the feature expression ability of the model and improve the accuracy of Re-ID.

Key words: Person Re-Identification(Re-ID), Generative Adversarial Network(GAN), feature extraction, multi-scale fusion, joint constraint

中图分类号: