【作者】 肖伟；

【导师】 李开；

【作者基本信息】 华中科技大学 ， 计算机应用技术， 2021， 硕士

【摘要】 随着视频大数据的日益积聚、硬件技术的不断突破和计算机视觉技术的迅猛发展,基于视频监控图像和深度学习的行人重识别技术在城市安防和智慧社区等领域中持续发力,应用前景广阔。然而,现实复杂场景下的行人重识别面临着非常严重的域偏差问题,光照、视角、分辨率等外部因素的差异都将对算法的精度产生严重干扰。因此研究行人重识别的域自适应问题对面向实际应用的算法落地有着极为重要的意义。基于对抗生成网络和聚类伪标签两种方法,本课题提出基于对抗生成网络的域自适应行人重识别算法CGSA-Re ID(Cross-domain Generation for Self-Adaption Re ID)和基于特征聚类和伪标签的无监督域自适应行人重识别算法CMUDA-Re ID(Contrastive Learning based on Hybrid Memory for Unsupervised Domain Adaptive Re ID),通过有效改进和独创设计实现了跨域行人重识别性能的大幅跃升。CGSA-Re ID包含三大功能模块:特征提取模块、跨域图像生成模块和域自适应模块,通过改进骨架网络和特征提取方法,获得更加鲁棒和更具辨识度的行人身份特征;通过设计基于共享图像对抗损失和域对抗损失的循环一致性自监督图像生成方法,在完整保留图像信息的前提下充分解耦行人身份特征并增强其域无关性,此外,通过引入目标域自训练方法进一步增强模型的域泛化能力。CMUDA-Re ID在前述改进骨架网络和特征提取方法的基础上,通过设计动态混合存储空间同时保留并持续更新源域类别质心、目标域聚类中心特征和目标域离群样本特征,为特征提取器的优化提供充分的监督信号;通过设计独特的聚类可靠性度量方法:聚类独立性度量和聚类紧凑型度量,保证监督信号的合理性和优化方向的正确性,进而增强了模型的跨域识别性能。多组跨域识别任务中的实验验证表明,本课题设计的两种方法均能有效缓解行人重识别的域偏差问题,识别精度均超过现有方法。更多还原

【Abstract】 With the increasing accumulation of video big data,continuous breakthroughs in hardware technology and rapid development of computer vision technology,person reidentification technology based on video surveillance images and deep learning continues to exert its strength in the fields of urban security and smart communities,and its application prospects are broad.However,person re-identification in complex realistic scenes faces a very serious problem of domain deviation,and the differences in external factors such as illumination,viewing angle,and resolution will seriously interfere with the accuracy of the algorithm.Therefore,studying the domain adaptation problem of person re-identification is extremely important for the implementation of practical application-oriented algorithms.Based on two methods of generative adversarial network and clustering pseudo-tags,this paper proposes a domain adaptive person re-identification algorithm based on generative adversarial network named CGSA-Re ID(Cross-domain Generation for SelfAdaption Re ID)and an unsupervised domain adaptive pedestrian re-recognition algorithm based on feature clustering and pseudo-labeling named CMUDA-Re ID(Contrastive Learning based on Hybrid Memory for Unsupervised Domain Adaptive Re ID),has achieved a significant jump in cross-domain person re-identification performance through effective improvements and original design.CGSA-Re ID includes three functional modules: feature extraction module,crossdomain image generation module and domain adaptation module.By improving the backbone network and feature extraction method,it can obtain more robust and recognizable pedestrian identity features;through design based on the cyclic consistency self-supervised image generation method based on the shared image adversarial loss and the domain adversarial loss,it fully decouples the identity features of pedestrians and enhances their domain independence on the premise of completely preserving the image information.In addition,the self-training method of the target domain is introduced to further enhance the domain generalization ability of the model.Based on the aforementioned improved backbone network and feature extraction method,CMUDA-Re ID retains and continuously updates the source domain category centroid,target domain cluster center features,and target domain outlier sample features by designing a dynamic hybrid storage space,and is a feature extractor The optimization provides sufficient supervision signals;by designing a unique cluster reliability measurement method: cluster independence measurement and cluster compactness measurement,the rationality of the supervision signal and the correctness of the optimization direction are ensured,thereby enhancing the span of the model domain recognition performance.Experiments in multiple sets of cross-domain recognition tasks verify that the two methods designed in this paper can effectively alleviate the domain bias problem of person re-identification,and the recognition accuracy exceeds the existing methods.更多还原