双主干伪装目标检测网络

Camouflaged object detection with dual backbone

针对伪装目标检测任务中存在检测精度有限的问题,通过引入双主干网络增加差异化信息,提出一种双主干伪装目标检测网络(Dual Backbone Network,DBNet).设计了双主干特征融合模块、边缘注意力模块和逐级细化模块.双主干特征融合模块将Res2Net50和PVT v2对原始图像提取的多级特征进行有效融合,获取丰富的全局上下文信息和局部上下文信息;边缘注意力模块根据生成的边缘预测图进一步计算边缘注意力图,使网络更加关注伪装目标的边缘细节;在逐级细化模块中,上一层的预测图和特征与当前层的特征依次经过粗预测细化结构和交叉查询注意力结构,该模块在标签监督下能够提供逐渐精确和细化的预测结果 .研究结果表明:在CAMO数据集,DBNet的S_(α)、F_(β)^(ω)和E_(φ)分别为0.877、0.838和0.932,MAE为0.042;在COD10K数据集,DBNet的MAE和E_(φ)分别为0.022和0.932;在NC4K数据集,F_(β)^(ω)和MAE分别为0.843和0.031.所提网络DBNet的检测性能优于其他23个伪装目标检测网络,所设计的3个模块能够有效提升网络对伪装目标的检测能力.

Addressing the issue of limited detection accuracy in camouflaged object detection tasks,this paper proposes a Dual Backbone Network(DBNet)by introducing a dual-backbone network to en-hance differentiation information.The DBNet comprises three main components:the dual backbone feature fusion module,edge attention module,and level-by-level refinement module.The dual back-bone feature fusion module effectively integrates the multi-level features extracted from the original im-age by Res2Net50 and PVT v2,thereby capturing rich global and local context information.The edge attention module calculates an edge attention map based on the generated edge prediction map,direct-ing the network’s focus towards the edge details of camouflaged objects.Within the level-by-level re-finement module,the prediction map of the previous layer and features are sequentially refined through coarse prediction refinement structures and cross-query attention structures,progressively improving prediction accuracy and refinement under label supervision.Experimental results on the CAMO data-ωset demonstrate DBNet’s performance with S_(α),F_(β)^(ω),and E_(ϕ)of DBNet being 0.877,0.838,and 0.932,respectively,and with an MAE of 0.042.On the COD10K dataset,DBNet achieves an MAE of 0.022 and E_(ϕ)of 0.932,while on the NC4K dataset,F_(β)^(ω)and MAE are 0.843 and 0.031,respectively.DBNet outperforms 23 other camouflaged object detection networks,with its three designed modules effectively enhancing the network’s detection capability.