融合注意力机制的弱监督迷彩伪装目标检测算法  

Weakly supervised camouflage object detection algorithm fused with attention mechanism

作　　者：杨辉 权冀川 梁新宇 郭安文 王中伟 Yang Hui;Quan Jichuan;Liang Xinyu;Guo Anwen;Wang Zhongwei(Command&Control Engineering College,Army Engineering University of PLA,Nanjing 210007,China;Unit 73658 of PLA,China)

机构地区：[1]陆军工程大学指挥控制工程学院,江苏南京210007 [2]中国人民解放军73658部队

出　　处：《网络安全与数据治理》2022年第9期81-91,共11页CYBER SECURITY AND DATA GOVERNANCE

摘　　要：随着计算机硬件和人工智能技术的发展,强监督目标检测算法已经取得了很大的成果。然而,强监督目标检测算法需要在大规模、标注精度高的数据集上进行训练。但在某些特定领域,上述条件要求过于苛刻。例如,军事上常用的迷彩伪装目标的图像数据集就比公共数据集更难获得且标注难度更大。因此,采用对数据集要求更低的弱监督目标检测算法来实现迷彩伪装目标的检测任务。由于图像中迷彩伪装目标与背景融合度较大,导致原始浅层特征感知伪监督目标定位(Shallow feature-aware Pseudo supervised Object Localization, SPOL)算法的检测精度相对较低。本文的核心是在SPOL算法的基础上融合注意力机制,通过加入注意力模块,让模型更加关注迷彩伪装目标的区域,以此来提高迷彩伪装目标的检测精度。With the development of computer hardware and artificial intelligence technology, strongly supervised object de-tection algorithms have achieved great results. However, strongly supervised object detection algorithms need to be trained on large-scale datasets with high annotation accuracy. But in some specific fields, the above conditions are too demand-ing. For example, image datasets of camouflage objects commonly used in military are more difficult to be obtained and labeled than public datasets. Therefore, this paper adopted a weakly supervised algorithm with lower requirements for datasets to detect the camouflage objects in images. Due to the high degree of fusion between the camouflage objects and the image background, the detection accuracy of the original SPOL( Shallow feature-aware Pseudo supervised Object Local-ization) algorithm was relatively low. The core of this paper was to integrate the attention mechanism into the SPOL al-gorithm. After added an attention module, the model could pay more attention to the area with camouflage objects. So,the detection accuracy of the camouflage objects was improved.

关 键 词：目标检测 弱监督算法 注意力机制 迷彩伪装目标 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]