Dual-Frequency Category-Guided Recalibration Network for Remote Sensing Image Semantic Segmentation

In remote sensing images, confusion between adjacent categories commonly occurs due to similar local textures across different land-cover types, where category boundaries often exhibit gradual transitions rather than clear separations. Furthermore, intra-class variability induced by complex environmental further weakens inter-class discriminability, making it difficult for conventional models to sufficiently distinguish adjacent categories.To order to tackle this challenge, this paper proposes the Dual-Frequency Category-Guided Recalibration Network (DFCGRNet), leveraging spatial and frequency domain modeling.The Dual-Frequency Interaction (DFI) module separates high- and low-frequency components via a customized wavelet transform, enhancing boundary details while preserving global structural coherence.Within the Frequency-Category Collaborative Enhancement (FCCE) module, the Wavelet Frequency Enhancement (WFE) unit adaptively strengthens boundary cues between adjacent categories through discrete wavelet modulation.The Category-Guided Semantic Recalibration (CGSR) module computes category centroid vectors for each category and constructs an inter-class cosine similarity matrix promote adjacent category distinction.Consequently, the DFCGRNet is empowered to enhance discriminative capacity between adjacent categories. Empirical evaluations show DFCGRNet achieves mIoU scores of 85.30%, 87.67%, and 55.11% on the Vaihingen, Potsdam, and LoveDA datasets, surpassing a range of prevailing approaches in segmentation accuracy. The complete implementation is available athttps://github.com/liumh12/DFCGRNet.