【摘要】 现有遥感图像分割方法在跨尺度特征融合时缺乏对低频结构与高频细节的协同处理，且无法对图像内容自适应响应，导致难以有效解决因自然变化、光照和阴影干扰引起的高分辨率遥感图像中类内差异大、类间差异小的问题。为此提出了一种自适应频率感知网络（AFANet）。首先，提出一种频率动态融合模块，通过自适应低通和高通滤波在保留低频结构的同时抑制高频噪音分量，并增强高频细节边界信息。其次，构建双域学习模块集成空间和频率信息，实现空间域局部细节与频域全局结构的联合建模。最后，引入一个细节增强模块，利用不同的差分卷积以增强模型的特征提取和泛化能力。在Vaihingen和Potsdam两个经典公开数据集上通过对比和消融实验的定量及可视化分析表明，AFANet在F1分数、总体精度和平均交并比等指标中优于7种先进的分割方法，验证了AFANet的优越性能。更多还原

【Abstract】 The existing remote sensing image segmentation methods lack the coordinated processing of low-frequency structures and high-frequency details during cross-scale feature fusion, and cannot adaptively respond to image content, making it difficult to effectively solve the problem of large intra-class variations and small interclass variations in high-resolution remote sensing images caused by natural changes, illumination, and shadow interference. To address these problems, an adaptive frequency-aware network(AFANet) is proposed and implemented. First, a frequency dynamic fusion module is proposed, which suppresses high-frequency noise components while retaining low-frequency structures through adaptive low-pass and high-pass filter, and enhancing high-frequency detail boundary information. Secondly, a dual-domain learning block is constructed to integrate spatial and frequency information to achieve joint modeling of local details in the spatial domain and global structures in the frequency domain. Finally, a detail-enhanced module is introduced to enhance the feature extraction and generalization capabilities of the model using different differential convolutions. The quantitative analysis and visualization results of the comparative and ablation experiments on the two classic public datasets,Vaihingen and Potsdam, show that AFANet outperforms seven state-of-the-art segmentation methods in terms of F1 score, OA and mIoU, with the mIoU reaching 85.13% and 87.81% respectively, verifying the superior performance of AFANet.更多还原