基于频率域可分离卷积的遥感图像道路分割方法OA北大核心CSTPCD

Segmentation of roads from remote sensing images is a challenging topic.Previously,most methods relied on convolutional neural networks,but these network models are difficult to capture long-distance feature information.The attention mechanism is known for its global vision,but it brings high computation burden.Convolution in frequency domain provides a novel mechanism to capture features over long-range,and an asymmetric convolution structure is introduced to achieve low computational cost.Based on it,this paper proposes a novel road segmentation network on remote sensing images,called lightweight separable Fourier filtered U-shape network(LSFU-Net).The LSFU-Net is composed of basic blocks for feature extraction in frequency domain and follows the pipeline of classical U-Net model.The separable complex convolution is used in basic blocks for feature extraction in frequency domain,which not only realizes the compression of model parameters,but also enhances the feature extraction ability of the model.Experimental results on the Massachusetts Roads Dataset and the DeepGlobe Road Dataset demonstrate that the LSFU-Net achieves excellent segmentation performance with fewer parameters.