基于多尺度上下文与边界生成的伪装物体分割OA

When segmenting camouflaging objects,the similarity between the foreground and background of the camouflaging objects leads to difficulties in feature extraction and boundary positioning.An innovative multi-scale context and boundary generation network is proposed.Not only global features are effectively extracted through multi-layer Transformer modules,but also necessary local detail information is retained.The proposed network combines of the context enhancement module and the cross-scale feature aggregation module can significantly enhance global con-text features and promote the efficient interaction of global context information.The introduction of the boundary gen-eration module can learn the boundary information in the features and accurately locate the boundaries of the camou-flaged objects.The experimental results show that,compared with other methods,the average maximum improvement of the four evaluation indicators of the four datasets by this network is 15.9 percentage points,13.4 percentage points,30.6 percentage points and 6.4 percentage points,respectively,verifying its better segmentation accuracy.