Addressing the issue of limited detection accuracy in camouflaged object detection tasks,thispaper 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-set demonstrate DBNet's performance with Sα,Fωβ,and Eϕ 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 Eϕ of 0.932,while on the NC4K dataset,Fωβ 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.
万方数据
维普
