摘要
伪装目标检测(camouflaged object detection,COD)是计算机视觉领域一项具有挑战性的基础研究.由于像素级注释的成本较高,研究者们通常采用涂鸦注释作为弱监督信号.然而,涂鸦标注存在信息过于稀疏且缺乏边缘信息等固有局限,这严重制约了模型的预测可靠性.针对这些问题,本文提出一种新颖的交互图推理网络(interactive graphical reasoning network,IGRNet),该网络通过图表示来推断伪装区域及其边缘之间的内在关系.具体而言,引入了图推理网络建模像素间的长距离依赖关系,设计了高效的图交互单元(graph interaction unit,GIU)增强异构特征的表征能力.同时,为提升模型的场景理解能力并充分利用不同特征间的互补性,构建了上下文增强模块(context enhancement module,CEM)实现多特征融合与上下文信息挖掘.此外,提出了自监督伪装检测损失(self-supervised camouflage detection loss,Lscd)来引导网络学习结构信息,进一步增强前景−背景的区分能力.在3个标准基准数据集上的大量实验表明,本文方法不仅显著优于现有弱监督算法,在某些评估指标上甚至超越了全监督方法的性能.
As a challenging fundamental research in computer vision area,camouflaged object detection(COD)usually take scribble annotations as weakly supervised signals due to the high cost of pixel-level annotations.To overcome the inherent limitations of scribble annotations with sparse information and lack of edge information,a novel Interactive Graphical Reasoning Network(IGRNet)was proposed in this paper,inferring the intrinsic relationships between the camouflaged regions and their edges with graph representations to improve the prediction reliability of models.Specifically,a graph inference network was introduced to model long-range dependencies between pixels,and an efficient Graph Interaction Unit(GIU)was designed to enhance the representation of heterogeneous features.Meanwhile,in order to improve the scene understanding ability of the model and make full use of the complementarity between different features,a Context Enhancement Module(CEM)was constructed to achieve multi-feature fusion and contextual information mining.In addition,a self-supervised camouflage detection loss(Lscd)was proposed to guide the network to learn structural information and further improve the foreground-background distinction ability.Extensive experiment results on three standard benchmark datasets show that the proposed method can not only significantly outperform existing weakly-supervised algorithms,but even surpass the performance of fully-supervised methods in some evaluation methods.
作者
张冬冬
王春平
付强
宋瑶
刘新海
ZHANG Dongdong;WANG Chunping;FU Qiang;SONG Yao;LIU Xinhai(Shijiazhuang Campus,Army Engineering University of PLA,Shijiazhuang,Hebei 050003,China;32368 Army,Beijing 100042,China)
出处
《北京理工大学学报》
北大核心
2025年第7期718-730,共13页
Transactions of Beijing Institute of Technology
基金
军内科研项目。
关键词
伪装目标检测
弱监督
涂鸦注释
图推理网络
上下文信息
camouflaged object detection
weak supervision
scribble annotation
graph inference networks
contextual information
