【作者】 张凯；

【导师】 葛宏伟；

【作者基本信息】 大连理工大学 ， 计算机科学与技术， 2020， 硕士

【摘要】 行人重识别的目标是在图像库中检索与查询图片具有相同身份并且由不同的相机拍摄的行人图片。行人重识别技术在智能监控、刑事侦查、智慧城市等众多领域发挥着重要作用。随着深度学习技术的发展,监督学习行人重识别模型在标注数据集上取得了令人满意的性能。但是监督学习方法需要大量的人工标注标签,在实际应用中代价过于昂贵。近年来,许多工作开始使用深度学习无监督域自适应迁移方法进行跨域的行人重识别研究。通过探索目标域潜在标签信息的标签估计方法实现了相对较好的效果,但与监督学习方法相比仍然有较大的差距。域自适应标签估计方法的关键是如何提高生成的伪标签的可靠性,进而训练一个适应目标域的行人重模型。为了生成更可靠的伪标签,本文提出判别学习网络LatentDLN来探索利用三种目标域潜在信息:行人图像的有效区域信息、最近邻居信息和相机风格信息。本文采用人体关键点检测技术提取行人图像的有效局部区域,基于有效区域信息设计多分支网络结构来学习行人图片的全局线索和局部线索,提高特征的判别能力。在距离度量过程中,本文使用基于最近邻居信息的重排序技术来提高排序的质量。为了利用相机风格信息,本文训练多域风格转换生成对抗网络为目标域每张图片生成相机风格转换的图片,然后将无监督聚类估计的标签传递给对应的风格转换图片来进行训练过程的数据增强。此外,为了处理无效区域导致的局部特征缺失情况,本文设计了启发式度量方法来度量不同行人图片之间的相似性。本文使用自训练机制训练提出的模型,即反复并交替地执行标签估计和模型训练直到模型收敛。为了增强初始模型在目标域中的特征表达能力,本文提出区域指导联合监督学习策略,进一步利用了目标域中潜在的有效区域信息和相机风格信息。在源域和目标域上联合训练的初始模型能够在一定程度上捕捉目标域中的全局局部线索和相机风格变化,为自训练过程的起始阶段提供更多准确的标签,进而提高了本文提出模型的最终性能。本文在大规模标准行人重识别数据集Market-1501、DukeMTMC-reID和MSMT17上进行了实验验证和分析,实验结果表明本文提出的方法在三个数据集上都取得了目前最好的性能,验证了本文方法在跨域行人重识别下的有效性。更多还原

【Abstract】 The person re-identification(Re-ID)task aims at retrieving the pedestrian images with the same identity as the query image but taken by different cameras in the gallery.Re-ID plays an important role in intelligent video surveillance,criminal investigation,smart city and many other fields.With the development of deep learning approaches,supervised Re-ID models have achieved satisfactory performance on the labeled benchmark dataset.However,the supervised methods require sufficient manually tagged labels,which is too expensive in the real-world applications.Recently,many works employ deep learning based unsupervised domain adaptation transferring learning methods to the study of cross domain Re-ID.The label estimation methods have achieved relatively superior results,but there is still a huge gap compared with the supervised learning method.The key of the domain adaptation label estimation approaches is how to improve the reliability of the generated pseudo labels,and then the Re-ID model is trained by using the pseudo labels to adapt to the target domain.To generate more reliable pseudo labels,this paper proposes a discriminative learning network with target domain latent information(LatentDLN)for exploring three types of latent information in the target domain,i.e.,the valid region information,the nearest neighbor information and the camera style information.First,this paper adopts key points detection for extracting the valid local regions of the person image,then a multi-branch network is designed to leverage the local and global cues,which improves the discriminative ability of feature representation.Second,we use the re-ranking mechanism based on the nearest neighbor information to improve the quality of the ranking list in the process of distance measurement.Third,to utilize the camera information,this paper trains a multi-domain image translation generative adversarial network to obtain camera style transferred images for each target image.The pseudo labels of true images estimated by the unsupervised clustering method are assigned to the corresponding generated images for data augmentation during the training process.Furthermore,to deal with the cases of feature missing caused by invalid regions,we propose a heuristic distance metric learning method to effectively evaluate the similarity between different images.Finally,the proposed LatentDLN is trained based on the self-training mechanism,i.e.,we repeatedly and alternatively conduct the label estimation and the training process until the model is stable.This paper proposes a region-guided jointly supervised learning strategy to improve the feature representation ability of the initial model in the target domain,which further utilizes the latent valid region and camera style information.The initial model jointly trained on the source and the target domain could to some extent capture the local and global cues and the camera variances in the target domain.It provides more accurate labels at the beginning phase of the self-training process and improves the final performance of the proposed model.In this paper,we conduct extensive experiments on three large-scale standard Re-ID datasets,i.e.,Market-1501,DukeMTMC-reID and MSMT17.Experimental results show that the proposed method significantly outperforms the state-of-the-art approaches,which demonstrates the effectiveness of the proposed method for cross-domain Re-ID.更多还原