CASIA OpenIR  > 毕业生  > 硕士学位论文
目标重识别中的鲁棒特征学习
凃鸣非
2022-05
Pages80
Subtype硕士
Abstract

     目标重识别是智能交通系统的关键和基础技术,是包括行人重识别和车辆 重识别在内的一项具有挑战性的任务。该任务的核心在于学习鲁棒高效的目标 视觉特征表达模型。其鲁棒性体现在两点,其一为同一数据分布下,模型对视 角变化、光线变化、姿态变化等外部干扰的鲁棒性。其二为不同数据分布下,模 型对分布差异(又称不同数据域之间的域偏移)的鲁棒性。针对第一个问题,本 文以深度神经网络为技术基础,增强了目标图片中存在的不同粒度特征之间的 交互,学习图片中目标的多层次特征,构建一个能够更全面地、更鲁棒地描述 图片特征的主干网路。针对第二个问题,本文以无监督域自适应任务中的对比 学习为方法基础,利用由簇类到样本层级的递进式对比学习,使用多源域数据 对目标域数据进行有指向性的特征级数据扩增,增强模型对域偏移的鲁棒性。

  论文的主要工作和创新点归纳如下:

1. 多粒度特征互学习目标重识别网络。目标重识别的核心任务是构建一 个可以提取精细的、具有判别力特征的主干网络,以缓解由视角变化和目标姿 态变化等因素导致的复杂的类内、类间差异难题。现有的方法专注于学习局部 的具有判别力的特征用于提升重识别模型的性能,这类方法通常只关注于图片 中的某一特定区域,多样性不足,并且在应对遮挡等特殊情况时不够鲁棒。为 了解决这个问题,本文提出了多粒度互学习目标重识别框架,该方法主要做出 了以下两点贡献:(1)本文将多粒度拼图打乱模块引入到目标重识别中,通过 破坏图片中原始存在的空间联系,驱使网络学习不同粒度的局部的有判别性特 征。(2)本文提出了一个无参数的多尺度特征重建模块,用于不同视觉粒度特 征之间的相互学习,使得无论是全局特征还是局部特征都具有更强的表征能力。 大量的实验证明了我们提出的模块的有效性以及我们的方法在行人和车辆重识 别基准上优于同期最先进的方法。

2. 基于梯度引导特征增强的无监督域自适应目标重识别。无监督域自适应 目标重识别旨在研究如何有效利用有标签的源域数据和无标签的目标域数据, 提高目标重识别模型在目标域上的性能。该任务的核心在于如何增强模型对域 间隙的鲁棒性,将模型学到的源域知识有效的迁移到目标域,同时充分利用无标签目标域数据之间隐藏的语义相似度关系,提高模型对目标域内类内类间差 异的鲁棒性。对此,本文提出了基于梯度引导特征增强的无监督域自适应目标 重识别框架。一方面利用多个源域生成多组身份有关、域无关的增强特征,通 过提高目标域数据的类内域分布多样性来增强模型对域间隙的鲁棒性。另一方 面,构建了一种双层对比学习框架,通过样本级和簇类级对比学习训练,有效 促进模型在目标域下的判别性鲁棒性特征学习。本文方法的有效性在多个行人 重识别和车辆重识别公开数据集上均得到了验证。

Other Abstract

Object re-identification is the key and basic technology of intelligent transportation system, and it is a challenging task including pedestrian re-identification and vehicle re-identification. It is designed to retrieve a given target object from images captured by dierent cameras. The core of this task is to extract robust image features through deep models. Its robustness is reflected in two points, one is the robustness against complex environmental factors such as viewing angle changes and light changes within the domain, and the other is the robustness against domain osets between multiple domains. In response to the first problem, this paper takes the metric learning deep neural network as the technical basis, studies the relationship between dierent granular features existing in the target image, learns the multi-level features of the target in the image, and builds a more comprehensive, A backbone network that more robustly describes image features. In response to the second problem, this paper is based on deep metric learning and self-supervised learning, using progressive comparative learning from cluster classes to sample level, and using multi-source domain data to perform directional feature-level data on target domain data. Augmentation to enhance the robustness of the model to domain shifts.

The main work and innovations of the paper are summarized as follows:

1. Multi-grain Feature Mutual Learning Object Re-identification Network.

The core task of object re-identification is to have a backbone network that can extract fine, discriminative features to handle complex intra- and inter-class transformations caused by viewpoint changes and object poses. Existing methods focus on learning local discriminative features to improve the performance of the re-identification model. Such methods usually only focus on a specific area in the image, have no diversity, and are not robust to special cases such as occlusion. To tackle this issue, this paper proposes the Multi-granularity Mutual Learning Network (MMNet) and makes two contributions. (1) We introduce the multi-granularity jigsaw puzzle module into object re-ID to impel the network to learn local discriminative features from multiple visual granularities by breaking spatial correlation in original images. (2) We propose a parameter-free multi- scale feature reconstruction module to facilitate mutual learning of features at multiple grain levels, thereby both global features and local features have strong representation capabilities. Extensive experiments demonstrate the eectiveness of our proposed module and that our method outperforms various state-of-the-art methods on pedestrian and vehicle re-identification benchmarks.

2. Unsupervised Domain Adaptive Object Re-Identification Based on Gradient- guidedFeatureAugmentation. Unsupervised domain adaptive object re-identification aims to study how to eectively utilize the labeled source domain data and unlabeled target domain data to improve the performance of the target re-identification model on the target domain. The core of this task is how to enhance the robustness of the model to the domain gap, eectively transfer the source domain knowledge learned by the model to the target domain, and make full use of the hidden semantic similarity relationship between unlabeled target domain data to improve the model. Robustness to intra-class and inter-class dierences within the target domain. In this regard, this paper proposes an unsupervised domain-adaptive object re-identification framework based on gradient-guided feature augmentation. From one perspective, multiple source domains are used to generate multiple sets of identity-related and domain-independent enhanced features, and the robustness of the model to domain gaps is enhanced by increasing the intra-class domain distribution diversity of the target domain data. Further, a two-layer contrastive learning framework is constructed, which eectively promotes the discriminative robust feature learning of the model in the target domain through sample-level and cluster-level contrastive learning and training. The eectiveness of our method has been verified on multiple public datasets for person re-ID and vehicle re-ID.

Keyword目标重识别 鲁棒特征学习 多粒度特征 多源域自适应
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/48602
Collection毕业生_硕士学位论文
Recommended Citation
GB/T 7714
凃鸣非. 目标重识别中的鲁棒特征学习[D]. 中国科学院自动化研究所. 中国科学院自动化研究所,2022.
Files in This Item:
File Name/Size DocType Version Access License
目标重识别中的鲁棒特征学习.pdf(8894KB)学位论文 限制开放CC BY-NC-SA
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[凃鸣非]'s Articles
Baidu academic
Similar articles in Baidu academic
[凃鸣非]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[凃鸣非]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.