Abstract: Aiming at the problems of large number of parameters, unstable gradient and low segmentation accuracy of the original SegNet network model, we proposed an improved model by constructing the SegNet combined with bottleneck block with residual, depthwise separable convolution and skip connection structure. The experimental results on aerial and satellite remote sensing image data sets show that the improved network model obtains better results in performance evaluation indexes such as accuracy, recall and F₁ value, which shows that the improved network model has good practical value in building segmentation task of remote sensing images.

Key words: residual, depthwise separable convolution, skip connection, remote sensing image, building segmentation