Computer Engineering and Applications ›› 2021, Vol. 57 ›› Issue (14): 187-193.DOI: 10.3778/j.issn.1002-8331.2004-0220

Previous Articles     Next Articles

Progressive Adversarial Learning for Weakly Supervised Object Localization

LUO Hanwu, LI Wenzhen, PAN Fucheng, JU Xiaoming   

  1. 1.Inner Mongolia Eastern Electric Power Company, State Grid Corporation of China, Hohhot 010010, China
    2.School of Software Engineering, East China Normal University, Shanghai 200062, China
  • Online:2021-07-15 Published:2021-07-14

基于渐进对抗学习的弱监督目标定位

罗汉武,李文震,潘富城,琚小明   

  1. 1.国网内蒙古东部电力有限公司,呼和浩特 010010
    2.华东师范大学 软件工程学院,上海 200062

Abstract:

Aiming to solve the problem of lacking fine location annotations in many datasets in practical applications, a weakly supervised object localization algorithm based progressive adversarial learning is proposed. Specifically, in order to solve the problem of training difficulties caused by dataset noise, self-paced learning is introduced to sort the training data according to the principle of simplicity to difficulty. In the network design, the weakly supervised object localization network is designed as a multi-label classification network, and the corresponding adversarial loss function is proposed to adapt to the task. Finally, in order to solve the problem that the existing methods often only pay attention to the most distinguishing part and can not locate the whole target, a pyramid adversarial erase mechanism is proposed to find the complete target in the final location map. The experimental results on several benchmarks show that the proposed algorithm can achieve object localization on datasets with high accuracy. It has certain competitiveness with the most advanced method.

Key words: progressive adversarial learning, weakly supervised localization, adversarial erase

摘要:

针对实际应用中大量数据集缺乏精细位置标注的问题,提出了一种基于渐进对抗学习的弱监督目标定位算法。具体来说,针对数据集噪声造成训练困难的问题,引入自步学习对训练数据按由简到难的原则进行排序。在网络设计上,将弱监督目标定位网络设计为多标签分类网络,并提出了相应的对抗损失函数适应目标定位任务。为了解决现有方法往往只关注最具辨别力的部分,无法定位整个目标的问题,提出一种金字塔对抗擦除机制以此在最后的定位图中发现完整的目标。在数个标准的数据集的实验表明,该算法具有较高的定位精度,与最先进的弱监督目标定位的方法相比具有一定的竞争力。

关键词: 渐进对抗学习, 弱监督定位, 对抗擦除