基于CBAM的弱监督目标检测  

Detection of Weakly-supervised Objects Based on CBAM

作　　者：刘均[1] 潘妍[1] 邓海航 LIU Jun;PAN Yan;DENC Hai-hang(School of Phys ics and Electroric Engineering,Northeast Petroleum Uriversity)

机构地区：[1]东北石油大学物理与电子工程学院

出　　处：《化工自动化及仪表》2025年第2期191-197,共7页Control and Instruments in Chemical Industry

基　　金：黑龙江省自然科学基金(批准号:LH2023A002)资助的课题。

摘　　要：由于实例级类别标注的缺失,弱监督目标检测网络在精确预测目标位置时面临显著挑战。当前主流策略倾向于采用分阶段学习,然而这一过程可能导致特定对象类别陷入局部最优。为克服此难题,提出一种新的端到端联合训练框架,即构建了一个集成多实例学习与边界框回归分支的统一网络架构,两者共享一个高效的主干网络以促进协同。同时,引入注意力机制于主干,深化特征中的位置信息挖掘。通过在基准数据集PASCAL VOC 2007、2012上的广泛实验验证,证实所提方法达到了较高的性能。Lacking of instance-level category labeling brings the weakly-supervised object detection network under challenges in accurately predicting target locations.The mainstream strategy at present time tends to adopt a staged learning strategy and this process leads a particular object category to being trapped in a local optimum.With the view to overcoming it,a new end-to-end joint training framework was proposed,which constructs the unified network architecture boating of the integrated multiple instance learning and bounding box regression,and efficient backbone network was shared to promote collaboration.At the same time,the attention mechanism was introduced into the backbone to deepen location information mining in the features.Extensive experiments on benchmark datasets PASCAL VOC 2007 and 2012 demonstrate that the proposed method achieves high performance.

关 键 词：弱监督目标检测 边界框回归 全监督检测 注意力机制 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]