Answers are grounded in this paper and cite their sources.
To tackle the challenges of insufficient global context modeling, blurred object boundaries, and inefficient cross-layer feature fusion in high-resolution urban remote sensing image semantic segmentation tasks, this paper proposes a hybrid Gabor-Fourier ResNet-Transformer Network (GF-ResFormer). The model is optimized for global-local modeling, edge feature enhancement, and multi-scale dynamic fusion. First, a Global Enhancement Module (GEM) is designed to model long-range dependencies across frequency spectra using the Fourier Transform and the Transformer, while fusing frequency-domain and spatial-domain features through a dynamic weighting mechanism, effectively enhancing global semantic consistency in complex scenes. Second, a multi-scale learnable Gabor Boundary Branch (GBB) is introduced, which utilises multi-scale Gabor filters to extract direction-sensitive texture features and adaptively strengthens edge responses through a boundary attention mechanism, significantly improving the boundary clarity of structured objects such as buildings and roads. A Transformer-Gabor Block (TGB) is proposed in the decoder segment, achieving balanced optimisation of global semantics and local details through a collaborative mechanism that integrates parallel global self-attention with local Gabor convolution. To address the problem of feature redundancy in skip connections during the training of remote sensing images, a multi-dimensional fusion strategy is developed, combining channel, spatial, and pixel-level attention mechanisms (CSP) to weigh the cross-layer features between the encoder and decoder, enhancing the segmentation accuracy of targets and heterogeneous regions. Experiments on the ISPRS Vaihingen, Potsdam, and LoveDA datasets show that GF-ResFormer attains mIoU scores of 68.71%, 74.88% and 66.30%, respectively, surpassing current state-of-the-art (SOTA) methods. Visualisation results demonstrate that this approach exhibits stronger robustness in segmenting complex targets, adhering to targets, and fragmented objects, offering a new perspective for precisely interpreting high-resolution remote sensing imagery. The code is available at the following link: https://github.com/YuchunHuang/GF_ResFormer.
No takes yet. Share an insight, caveat, or question.
Cao et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: