多视野精细分析下的弱监督目标定位算法  

Weakly supervised object localization based on multi-view fine analysis

作　　者：张英俊[1] 贾聪聪 谢斌红[1] ZHANG Ying-jun;JIA Cong-cong;XIE Bin-hong(School of Computer Science and Technology,Taiyuan University of Science and Technology,Taiyuan 030024,China)

机构地区：[1]太原科技大学计算机科学与技术学院,山西太原030024

出　　处：《计算机工程与设计》2024年第6期1750-1756,共7页Computer Engineering and Design

基　　金：山西省重点研发计划基金项目(201803D121048、201803D121055);山西省基础研究计划基金项目(20210302123216)。

摘　　要：针对多尺度目标定位精度较差,难以捕获完整目标边界的问题,设计一种多视野精细分析模块并融入通道与空间注意力机制抑制背景噪声的干扰,获取多尺度目标的高分辨率特征。利用随机特征选取模块获取特征图随机位置的组合,聚合多个位置图获取最具辨别性的位置及其它位置的信息,融合浅层生成的类激活图与聚合类激活图获取细粒度位置信息,捕获完整的目标边界。与现有的弱监督定位方法相比,在解决多尺度目标定位效果差和局部最优问题上具有一定的优势。Aiming at the problems of poor positioning accuracy of multi-scale objects and difficulty in capturing the complete object boundary,a multi-field fine analysis module was designed and integrated with channel and spatial attention mechanism to suppress the interference of background noise,to obtain the high-resolution features of multi-scale objects.The random feature selection module was used to obtain the combination of random positions of the feature map,and multiple location maps were aggregated to obtain the most discriminative location and other location information.The shallow-generated class activation map and the aggregate class activation map were fused to obtain the fine-grained location information and capture the complete object boundary.Compared with the existing weakly supervised localization methods,it has certain advantages in solving the problem of poor localization effect and local optimization of multi-scale objects.

关 键 词：弱监督学习 目标定位 多尺度特征融合 注意力机制 全局平均池化 类激活图 正则化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]