Abstract:

Deep learning segmentation method is one of the hot topics in the field of remote sensing image segmentation. The mainstream deep learning methods include convolutional neural networks, transformer neural networks, and a combination of the two. Feature extraction is an important part of image segmentation. In addition to using convolution and other methods to extract features, recent research has focused on some new feature extraction paradigms, such as graph convolution and wavelet transform. In this article, the region construction attribute of clustering algorithms is utilized, and the improved clustering algorithm is used as the backbone feature extraction module while the convolution and visual transformer are used as auxiliary modules to obtain richer feature representations. On the basis of the module, a new hierarchical remote sensing image semantic segmentation network (NHNet) is proposed. The performance of NHNet semantic segmentation is evaluated and compared with other methods on the LoveDA remote sensing dataset. The results show that NHNet based on multi-feature extraction achieved competitive performance, with an average intersection-to-union ratio of 49.64% and a score of 65.7%. At the same time, ablation experiments show that the auxiliary module improves the accuracy of clustering algorithm segmentation, increasing the average intersection-to-union ratio of NHNet by 1.03% and 2.41%, respectively.

Key words: remote sensing images, semantic segmentation, clustering algorithm, convolutional neural network, self attention