Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (17): 132-142.DOI: 10.3778/j.issn.1002-8331.2204-0123

• Pattern Recognition and Artificial Intelligence • Previous Articles     Next Articles

Weakly Supervised Object Detection Algorithm Based on Adaptive Instance Optimization

LIU Zhoufeng, WANG Kaihua, TIAN Bo, LI Chunlei   

  1. School of Electrical and Information Engineering, Zhongyuan University of Technology, Zhengzhou 450007, China
  • Online:2023-09-01 Published:2023-09-01

基于自适应实例优化的弱监督目标检测算法

刘洲峰,王凯华,田博,李春雷   

  1. 中原工学院 电子信息学院,郑州 450007

Abstract: Aiming at the local optimization problem in the existing weakly supervised object detection algorithms, a weakly supervised object detection algorithm based on context-aware and adaptive multiple instance refinement is proposed in this paper. A global contextual attention module is designed to activate the whole features of the target object and alleviate the part domination problem. A parametric spatial dropout module is proposed to block the most discriminative region by utilizing the generate mask for preventing the detector taking the most discriminative region as the target object, which can conductive to jump out of the local optimal solution and meanwhile select the better candidate boxes with more complete target object. An adaptive supervised aggregation function is introduced to dynamically change the aggregation criteria, and then the effective supervision for training the online refinement branch is mined, which will improve the detection performance after several iterations. The experimental results illustrate that the detection accuracy on the Pascal VOC 2007 and VOC 2012 datasets is 52.1% and 46.6%, respectively; the localization accuracy is 68.1% and 66.4%, respectively, which are better than the existing detection algorithms and effectively solve the local optimal problem.

Key words: local saliency, context awareness, adaptive multiple instance refinement, adaptive supervised aggregation function

摘要: 针对现有弱监督目标检测算法存在局部最优问题,提出一种基于上下文感知和自适应多实例细化相结合的弱监督目标检测算法。设计全局上下文注意力模块激活目标对象的整体特征,缓解局部区域支配问题;提出参数化空间丢失模块,利用生成掩模遮挡最具辨别力的区域,避免检测器将其所在的候选区域作为目标对象,从而使模型跳出局部最优解,以便选出包含更完整目标对象的候选框;通过引入自适应监督聚合函数,动态改变聚合标准,挖掘出用于训练在线细化分支的有效监督,并经过多次迭代优化提升检测性能。实验结果表明,在Pascal VOC 2007和VOC 2012数据集上检测精度分别为52.1%、46.6%;定位精度分别为68.1%、66.4%,优于已有检测算法,且有效解决了局部最优问题。

关键词: 局部显著, 上下文感知, 自适应多实例细化, 自适应监督聚合函数