边缘信息引导的伪装目标检测
Edge information guided camouflaged object detection
吴涛林 1葛斌 2钱亚红 3夏晨星 2裴佳佳1
作者信息
- 1. 安徽理工大学计算机科学与工程学院,安徽淮南 232001
- 2. 安徽理工大学计算机科学与工程学院,安徽淮南 232001;合肥综合性国家科学中心能源研究院,安徽合肥 230031
- 3. 河南中烟工业有限责任公司安阳卷烟厂,河南安阳 455004
- 折叠
摘要
伪装目标检测(camouflaged object detection,COD)旨在检测隐藏在复杂背景中的伪装目标.由于伪装目标的特点:前景与背景纹理相似、边缘对比度低,导致现有方法得到的预测图像边缘模糊、小目标区域缺失.因此,本文提出了边缘信息引导的伪装目标检测网络(edge information guided network,EIGNet).首先,通过低层特征和高层特征对目标的边缘进行显式建模,充分提取目标的边缘特征指导后续特征表示.然后,通过双分支结构处理不同维度的伪装目标.其中,全局分支用以提取全局上下文信息强调大目标的全局贡献,局部分支用以挖掘丰富的局部低级线索增强小目标的特征表示.最后,采用自顶向下的方式实现相邻层特征的逐步融合,得到具有精细边缘和完整区域的预测图像.在3个伪装数据集上的实验结果表明本文方法优于其他15个模型,在NC4K数据集上平均绝对误差(mean absolute error,MAE)降至0.044.
Abstract
Camouflaged object detection(COD)aims to detect camouflaged objects hidden in complex backgrounds.Due to the characteristics of camouflaged object,such as similar foreground and background textures and low-contrast edges,existing methods often produce blurry edge of prediction image and miss small object regions.Therefore,this paper proposes an edge information guided network(EIGNet).First,the edge of the camouflaged object is explicitly modeled through low-level and high-level features,which fully extract the edge features of the object to guide subsequent feature representations.Then,a dual-branch structure is used to process different dimensions of camouflaged object.The global branch is used to extract global contextual information to emphasize the global contribution of large objects,while the local branch is used to mine rich local low-level clues to enhance the feature representation of small objects.Finally,a top-down manner is used to gradually aggregate adjacent layer features to obtain a prediction image with fine edges and complete regions.Experimental results on three camouflaged datasets show that our method outperforms 15 other models,with a mean absolute error(MAE)of 0.044 on the NC4K dataset.
关键词
伪装目标检测(COD)/边缘语义/特征增强/特征聚合Key words
camouflaged object detection(COD)/edge semantics/feature enhancement/feature aggrega-tion引用本文复制引用
基金项目
国家自然科学基金(62102003)
国家重大专项(2020YFB1314103)
安徽省自然科学基金(2108085QF258)
安徽省博士后基金(2022B623)
安徽省高等学校自然科学研究项目(KJ2020A0299)
出版年
2024
NETL
NSTL
万方数据