摘要
现有行人搜索方法着重于从有限的标注场景图中学习有效的行人表征,虽然这些方法取得了一定的效果,但学习更具有身份辨别力的行人表征通常依赖于大规模的标注数据,而获取大规模的标注数据是一个资源、劳动密集型的过程。为此,该文提出了一种场景图多粒度遮挡特征增强算法,对原始场景图进行多粒度随机遮挡,扩充训练数据,并从遮挡后的场景图中生成具有多样化信息的虚拟特征,最后利用生成的虚拟特征增强真实特征中的行人表征。进一步,基于生成对抗学习,该文设计了多粒度特征对齐模块,用于对齐遮挡图像特征和原始图像特征,保持两者语义一致性。实验结果表明,在CUHK-SYSU和PRW数据集上,该算法能够显著提升行人搜索任务的搜索精度。
Existing person search methods focus on efficiently learning pedestrian representations from limited labeled scene images.Although these methods have achieved good results,learning more identity-discriminative pedestrian representations usually relies on large-scale labeled images,while obtaining large-scale labeled data is a resource and labor intensive process.Therefore,we propose a novel multi-granularity occlusion feature enhancement algorithm for person search,which first performs multi-granularity random occlusion on original scene images to expand the training data,and then generates virtual features with diverse information from the occluded scene images.Finally,the generated virtual features are used to enhance the pedestrian representation in the real features.Furthermore,based on generative adversarial learning,a multi-granularity feature alignment module is designed to align the occluded image features and the original image features,and thereby maintain their semantic consistency.Experiments on CUHK-SYSU and PRW datasets show that the proposed algorithm can significantly improve the search accuracy of person search.
作者
苗春玲
张红云
吴卓嘉
张齐贤
苗夺谦
MIAO Chunling;ZHANG Hongyun;WU Zhuojia;ZHANG Qixian;MIAO Duoqian(College of Electronic and Information Engineering,Tongji University,Shanghai 201804,China;Key Laboratory of Embedded System and Service Computing Ministry of Education,Tongji University,Shanghai 201804,China)
出处
《智能系统学报》
北大核心
2025年第1期230-242,共13页
CAAI Transactions on Intelligent Systems
基金
国家重点研发计划项目(2022YFB3104700)
国家自然科学基金项目(62376198,62163016).
关键词
深度学习
计算机视觉
行人搜索
目标检测
粒计算
数据处理
特征提取
生成对抗网络
对齐
deep learning
computer vision
person search
object detection
granular computing
data processing
feature extraction
generative adversarial networks
alignment
