Building Image Segmentation Method with Multi-Attention Mechanism

In order to solve the problem of inaccurate edge segmentation and loss small buildings caused by UNet which is difficult take into account both global features local features, CSUNet proposed based on coordinate attention self-attention. The fuses in encoder, designs a Double-channel Skip Connection Transformer (DSCT) model skip connection, feature fusion module (FFM) CBAM channel fuse output connection with upsampling result decoder. tested instance dataset typical Chinese cities WHU East Asia satellite dataset. On cities, PA reaches 0.9390 IoU 0.8227, dataset, 0.9847 0.8332. Compared UNet, all indicators are improved. Visually, can more accurately extract building details such as edges corners, location contour buildings. Experiments show that improve performance extraction.