Answers are grounded in this paper and cite their sources.
Surveying building damage in post-disaster image segmentation is a difficult problem due to noise, low contrast, and ambiguous boundaries in real-life clinical images. The existing transformer models face problems such as feature redundancy, semantic inconsistencies, and poor boundary reconstructions. To overcome such drawbacks, this study introduces a new model named Reconstruction Spatial-Pixel Fusion Multi-Scale Transformer (RSPF-MST). Specifically, the proposed architecture has three main novelties: (i) Reconstruction Spatial-Pixel Convolution (RSPC) is designed to remove redundant features with the help of spatial and channel-wise reconstructions; (ii) Channel Interaction Fusion (CIF) is designed to strengthen multi-scale semantic consistency with cross-channel attention; and (iii) Pixel-Aware Up-sampling Guidance (PAUG) aims at improving boundary accuracy by introducing pixel-level dynamic weights. Moreover, a Dynamic Receptive Multi-Scale Context (DRMSC) module is added in the decoder for capturing context information at various levels. A pruning strategy is adopted for efficiency. Experimental analysis demonstrates that the proposed model significantly improves segmentation accuracy, achieving 98.43% and 98.53%, and IoU scores of 96.13% and 96.12% on the Ida-BD and RescueNet datasets, respectively, while maintaining robustness, particularly for low-quality post-disaster images.
No takes yet. Share an insight, caveat, or question.
Reddy et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: