Edge Refinement Detection of Camouflage Targets Based on Deep Texture Features
To solve the problem of incomplete spatial information and fuzzy target boundaries in conventional Camouflage Object Detection(COD),a COD algorithm based on depth texture features and edge thinning is proposed.Based on the texture difference and edge details of the target,the algorithm designs Context Texture Difference Amplification Module(CTDAM),Feature Boundary Search Module(FBSM),and Boundary Inference Module(BIM).CTDAM uses global receptive field coverage and parallel multi-branch hybrid convolution to highlight the texture differences of occluded camouflage targets.Additionally,it introduces local attention and position channel perception to guide feature cross-layer fusion in Attention Feature Fusion Module(AFFM).Therefore,it achieves the effect of balancing local information and enhancing semantic information of global context.FBSM uses a self-attention mechanism to combine low-and high-level features to deal with the dependence between different boundary pixels;and BIM uses the boundary guidance factor provided by FBSM to guide the fusion features to infer the real target and refine the edge details.Quantitative and qualitative experiments are conducted on the CAMO,CHAMELEON,and COD 10K datasets using four objective evaluation indices.The results demonstrate that the proposed algorithm is superior to other eight algorithms.On the COD 10K dataset,the Mean Absolute Error(MAE)is 0.034.
国家科技期刊平台
NSTL
万方数据
