High-resolution remote sensing images suffer from large intra-class variance, high inter-class similarity, and significant scale variations, leading to incomplete segmentation and imprecise boundaries. To address these challenges, Transformer-based methods, despite their strong global modeling capability, often suffer from feature confusion, weak detail representation, and high computational cost. Moreover, existing multi-scale fusion mechanisms are prone to semantic misalignment across levels, hindering effective information integration and reducing boundary clarity. To address these issues, a Category-Guided Transformer (CIGFormer) is proposed. Specifically, the Category-Information-Guided Transformer Module (CIGTM) integrates global and local branches: the global branch combines window-based self-attention (WSAM) and window adaptive pooling self-attention (WAPSAM), using class predictions to enhance global context modeling and reduce intra-class and inter-class confusion; the local branch extracts multi-scale structural features to refine semantic representation and boundaries. In addition, an Adaptive Wavelet Fusion Module (AWFM) is designed, which leverages wavelet decomposition and channel-spatial joint attention for dynamic multi-scale fusion while preserving structural details. Extensive experiments on the ISPRS Vaihingen and Potsdam datasets demonstrate that CIGFormer, with only 21.50 M parameters, achieves outstanding performance in small object recognition, boundary refinement, and complex scene parsing, showing strong potential for practical applications.
Ni et al. (Tue,) studied this question.