MFENet: Multi-Feature Extraction Net for Remote Sensing Semantic Segmentation

In this paper, we tackle the remote sensing semantic segmentation task by capturing feature information across multiple scales, all channels, and global locations. Different from previous works that simply use U-net to extract multi-scale features, we further improve U-net and propose a Multi-Feature Extraction Network (MFE-Unet). Specifically, we propose the MFE module, which uses both dilated convolution module and two attention modules. Dilated convolution is used to enhance U-net’s ability to represent multi-scale information. The two attention modules refer to the channel attention module and the pixel attention module. Channel attention maps all channels centrally, assigns weights uniformly, and adaptively adjusts the importance of each channel’s information. Pixel attention treats features at each location as the same individual, and similar features will be associated together to further improve feature representation. We conducted multiple sets of experiments on the "AI+" remote sensing image dataset. Experiments show that our network is sufficient against several advanced models.