一种基于区域权重平滑的弱监督目标定位方法  

A Weakly Supervised Object Localization Method Based on Region-weight Smoothing

作　　者：刘富[1] 罗冰[1] 裴峥[1] LIU Fu;LUO Bing;PEI Zheng(Center for Radio Administration&Technology Development,Xihua University,Chengdu 610039 China)

机构地区：[1]西华大学无线电管理技术研究中心,四川成都610039

出　　处：《西华大学学报（自然科学版）》2022年第3期1-7,共7页Journal of Xihua University:Natural Science Edition

基　　金：国家自然科学基金青年项目(61801398);西华大学“青年学者后备人才”支持计划项目;西华大学汽车测控与安全四川省重点实验室(QCCK2019-005)。

摘　　要：由于图像类别标签的弱监督目标定位方法存在定位区域仅覆盖目标最具有显著性类别信息部位的问题,同时,区域的类别信息响应受到关键权重的影响,且关键权重的不均衡导致了定位区域响应的稀疏性,因此,提出一种基于区域权重平滑的弱监督目标定位方法。文章设计了自适应标准差正则项,以缩小关键权重差异,从而在保留网络分类能力的同时平滑定位区域。在多个数据集上实验的结果表明,采用该方法所得的定位区域覆盖面更广,定位精度更高。Weakly supervised object localization methods with category supervision suffer from the problem that they tend to merely cover the most discriminative components of the object.And the category response of the region is affected by key weights,and the imbalance of them leads to the sparsity of object location.So this paper proposed a solution based on region-weight smoothing.This paper designed an adaptive standard deviation regularization to shrink weights discrepancy,which could smooth object location while preserving the classification performance.Results of experiments on several datasets show that this method could generate a wider area and achieve higher accuracy.

关 键 词：弱监督目标定位 正则化 类别激活谱 图像分类 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TP181[自动化与计算机技术—计算机科学与技术]