摘要
伪装目标检测(Camouflaged Object Detection,COD)旨在识别视觉上与周围环境高度融合的隐蔽目标。针对现有COD方法存在的边界特征建模不准确以及多特征融合效率低下等问题,提出了一种基于边界感知和自适应融合的渐进式伪装目标检测网络(Boundary-sensitive Adaptive Fusion Progressive Network,BAFPNet)。具体来说,该网络首先利用主干网络提取多层初始特征,继而通过DenseASPP(Densely connected ASPP)模块增强初始特征,并提出多尺度边界提取模块(Multi-scale Boundary Extraction Block,MBEB),利用非对称卷积与空间注意力机制,从初始特征中的浅层细节特征与深层语义特征中捕获精确边界信息;然后提出双注意力门控融合模块(Dual-Attention Gated Fusion Module,DAGFM)通过通道-空间双注意力机制和门控单元实现多尺度特征和边界信息的动态自适应融合;最后采用基于U形残差的深度可分离残差细化模块(Depthwise Residual Refinement Block,DRRB)实现预测结果的精细化重建。在3个基准数据集上的实验表明,本方法在4个评价指标上均显著优于现有方法,其中结构相似性指标(S-measure)提升尤为显著。
Camouflage object detection(COD)focuses on identifying concealed targets that are visually indistinguishable from their surrounding environment.To address the issues of inaccurate boundary feature modeling and low efficiency in multi-feature fusion in existing COD methods,this paper proposes a progressive camouflage object detection network,termed boundary perception and adaptive fusion(BAFPNet).Specifically,the backbone is utilized to extract multi-level initial features firstly.Then,the DenseASPP module is employed to enhance the initial features,and a multi-scale boundary extraction block(MBEB)is proposed,which captures precise boundary information from the shallow detail features and deep semantic features in the initial features by using asymmetric convolution and spatial attention mechanism.Subsequently,a dual-attention gated fusion module(DAGFM)is proposed to achieve dynamic adaptive fusion of multi-scale features and boundary information through channel-spatial dual-attention mechanism and gated units.Finally,a depthwise residual refinement block(DRRB)based on U-shaped residual is adopted to achieve refined reconstruction of the prediction results.Experiments on three benchmark datasets show that this method significantly outperforms existing methods in four evaluation metrics,with a particularly significant improvement in the structural similarity index(S-measure).
作者
黄祖卿
潘晴
田妮莉
HUANG Zuqing;PAN Qing;TIAN Nili(School of Information Engineering,Guangdong University of Technology,Guangzhou 510006,China)
出处
《激光杂志》
北大核心
2025年第10期48-55,共8页
Laser Journal
基金
国家自然科学基金(No.61901123)。
关键词
伪装目标检测
自适应融合
注意力机制
边界感知
camouflage object detection
adaptive fusion
attention mechanism
boundary perception
