基于可扩展生成对抗网络的跨域跨相机行人重识别  

Cross Domain and Cross Camera Person Re-identification based on Scalable Generative Adversarial Network

作　　者：沈茜[1] 何福男[1] SHEN Qian;HE Fu-nan(College of Artificial Intelligence,Suzhou Vocational Institute of Industrial Technology,Jangsu,Suzhou 215104)

机构地区：[1]苏州工业职业技术学院人工智能学院,江苏苏州215104

出　　处：《贵阳学院学报(自然科学版)》2025年第1期92-98,111,共8页Journal of Guiyang University:Natural Sciences

基　　金：2023年江苏省高等教育教改研究重点课题(2023JSJG798)。

摘　　要：为改善行人重识别模型在跨域和跨相机场景下性能大幅下降的问题,提出了跨域跨相机的行人重识别框架,结合了可扩展生成式对抗网络(S-GAN)的图像风格迁移,和基于标签判别嵌入向量(IDE)的重识别卷积神经网络(CNN)模型。所提S-GAN利用循环一致性损失解决了多相机风格迁移中目标域数据无标注问题,利用ID映射损失确保合成图像的行人ID不变性,并通过语义一致性损失在跨相机和跨域风格迁移中保留关键语义信息(行人前景信息)。此外,利用标签平滑归一化(LSR)技术解决合成图像噪声问题。两个大规模公开数据集上的实验结果表明,使用所提S-GAN进行跨相机和跨域图像风格迁移后得到的合成图像质量显著优于广泛使用的CycleGAN方法,且所提行人重识别框架在半监督(同域跨相机)和无监督(跨域)场景下取得了优于其他先进方法的性能。In order to improve the problem that the performance of person re-identification models is greatly degraded in cross-domain and cross-camera settings,a cross-domain and cross-camera person re-identification framework is proposed,which combines the image style transfer with Scalable Generative Adversarial Networks(S-GAN),and re-identification convolutional neural network(CNN)models based on label discriminative embedding(IDE).In the proposed S-GAN,the cycle consistency loss is used to realize unsupervised image style transfer among multiple cameras,the ID mapping loss is used to ensure the pedestrian ID invariance in the reconstructed images,and semantic consistency loss is utilized to preserve key semantic information(foreground information)during cross-camera and cross-domain image style transfer.In addition,Label Smoothing Regularization(LSR)technique is used to solve the noise problem introduced by transferred images.Experimental results on two large-scale public datasets show that the quality of generated images obtained after cross-camera and cross-domain image style transfer using the proposed S-GAN is significantly better than the widely adopted CycleGAN method,and the proposed re-identification framework outperforms other state-of-the-art methods in semi-supervised(same-domain cross-camera)and unsupervised(cross-domain)scenarios.

关 键 词：行人重识别 生成式对抗网络 卷积神经网络 标签判别嵌入向量 循环一致性 标签平滑归一化 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]