关键词: 矿区建筑物提取, 语义分割, SSC-SeNet, 注意力机制, 坐标卷积, 卷积神经网络, 特征融合

Abstract:

The U-Net encoder-decoder network structure is used to partition most buildings in mining areas, but the encoder-decoder structure does not make full use of the semantic and spatial features, resulting in low segmentation accuracy. Aiming at the defects of existing building extraction methods, a semantic spatial consistency semantic segmentation network (SSC-SeNet) is proposed. Firstly, the network uses a multi-channel structure to extract and integrate semantic features, spatial features, and consistency features. Secondly, a space extraction channel is introduced in the first three coordinate convolution of the main channel, and a Gabor Fourier filter is designed for further extraction of spatial features. Then, a semantic extraction channel is introduced at each layer of conventional convolution blocks in the main channel to improve the capability of semantic feature extraction. Finally, the feature fusion module is used to fuse the features of spatial extraction channel, semantic extraction channel and main channel, and the final segmentation image is generated. Experiments on the building data set of Xiangtan manganese mine with a resolution of 0.03 m show that the crossover ratio of SSC-SeNet is as high as 88.47% and the overall accuracy is 97.09%, both of which are ahead of mainstream traditional networks such as U-Net, and overfitting problems are overcome due to its lightweight characteristics.

Key words: mining building extraction, semantic segmentation, SSC-SeNet, attention mechanism, coordinate convolution, convolutional neural network, feature fusion