Multi-scale Wavelet Frequency Channel Attention for Remote Sensing Image Segmentation

Among recent developments in semantic segmentation, deep convolutional encoder-decoder has become the main-scheme model for remote sensing images. In this paper, we propose a architecture similar to U-Net for remote sensing image segmentation that uses wavelet frequency channel attention (WFCA) blocks as the attention mechanism to extract rich semantic features, which not only contain local information in spatial domain, but also consider frequency details in frequency domain. Then we fuse WFCA blocks with multi-scale skip connections to become multi-scale wavelet frequency channel attention (ms-WFCA) blocks for better utilizing features from different scales. Finally, the proposed method shows promising results on the Potsdam dataset.