FCMF-UNetFormer: A Full-Connected Multi-Scale Fused UNetFormer for Urban Scene Parsing

With the rapid growth of high-resolution remote sensing imagery, fine-grained urban scene parsing faces increasing challenges such as diverse object categories, large scale variations, and complex boundaries. Towards this issue, traditional CNNs are limited by local receptive fields; while Transformers which are, strong in global modeling struggle with hierarchical feature fusion. In this paper, we propose the FCMF-UNetFormer for urban scene parsing, a fully connected multi-scale fusion network which solves the semantic gap between shallow and deep features of transformer. In specific, we introduce a fully connected skip structure to enable progressive multi-scale interaction, a MultiSkipFusion module for continuous deep-shallow semantic integration, and a Proportional Skip Fusion (PSF) module that adaptively aligns and fuses features through learnable proportional weights. Experiments on the Vaihingen and LoveDA datasets demonstrate that the FCMF-UNetFormer achieves notable improvements over its baseline (i.e. UNetFormer) and outperforms recent state-of-the-art models in mIoU, F1-score, boundary quality, and overall accuracy, showing strong robustness and generalization.