HR_BMNet: a boundary perception and multiresolution fusion network for semantic segmentation of remote sensing

The conventional approach for semantic segmentation of remote sensing imagery using encoder-decoder convolutional neural networks relies on the output of prior feature maps sequentially without considering the interactions between neighboring contextual feature maps with multiple resolutions. While the standard HRNet proposal has successfully improved multi-resolution semantic and spatial features to address the aforementioned issues, its lack of emphasis on boundary perception often results in inadequate target segmentation. Furthermore, a frequent occurrence of multiresolution contextual interaction in HRNet leads to the addition of a significant quantity of redundant information and amplifies the complexity of the model. Hence, to tackle the abovementioned issues, we propose a semantic segmentation network identified as HR-BMNet, which incorporates boundary sensitivity and multiple-resolution learning. The idea associated with standard HRNet is adopted as the foundational architecture. We extend novel boundary perception and multi-resolution fusion attention modules, integrating channel attention mechanisms. The strategy provides an ex-tensive optimization of edges and the efficient capture of crucial multi-scale features. During the feature combination stage, the boundary insights are employed to augment the semantic information, thereby mitigating the spatial details loss, enhancing the intra-class semantic consistency, and achieving superior segmentation. The efficacy of the proposed method is validated through comparison and ablation experiments conducted on the ISPRS Vaihingen and CSRSD datasets. Among the experiments conducted, the best ones attained a mean Intersection over Union (mIoU) of 72.11% on the Vaihingen dataset and 89.28% on the CSRSD dataset, respectively.