Feature Fusion Network with Parallel Structure for Building Segmentation
Remote sensing building segmentation refers to pixel-level segmentation of buildings in remote sensing images,including precise outlines of buildings and detailed internal information.However,the unique characteristics of remote sensing images,with shadows similar in color to that of buildings,often result in under-segmentation,whereas factors such as tree occlusion can easily result in over-segmentation.A Multi-Feature Fusion Network(MFF-Net)based on a parallel structure is presented to address weak shadow interference,visible jagged edges of segmentation,and poor segmentation of building outlines in remote-sensing photos.The decoder uses ResNet-50 as its backbone and numerous parallel-structured dual-channel mask branches to reconstruct feature maps at various scales.To strengthen the critical edge features,an improved Convolutional Block Attention Module(CBAM)is further added to each branch structure.The bidirectional channel mask competition module is subsequently used to adjust channel interaction,thereby completing feature fusion.Experimental results on the ISPRS Potsdam and ISPRS Vaihingen datasets show that compared with existing mainstream segmentation networks,the global accuracy,precision,recall,F1 value,and mean Intersection over Union(mIoU)are improved to varying degrees.The precision of the Vaihingen dataset is 96.22％,the F1 value 95.55％,and mIoU 92.16％.On the Potsdam dataset,the precision is 96.95％,the F1 value 96.32％,and mIoU 93.40％.The extracted building contours are complete and clear,and resulted in strong interference resistance.
万方数据
维普
