基于频域特征的人群计数单域泛化

Single Domain Generalization Based on Frequency Domain for Crowd Counting

摘要: 在跨域人群计数任务中，源域与目标域之间的特征分布差异会导致模型泛化性能显著下降。为此， 本文提出了一种基于频域特征解耦的单域泛化人群计数方法。具体而言，我们首先利用二维傅里 叶变换对输入图像进行频域分解，显式分离出高频与低频成分。其中，高频特征包含与域无关的 边缘与纹理信息，而低频特征则反映了图像的全局结构和密度分布。考虑到密度回归中局部密度 变化平缓且主要受低频成分影响，我们引用了内容错误掩码以过滤低频特征中的域特异信息，并 构建高频引导的空间注意力机制，实现频域特征的有效融合。通过注意力一致性约束，进一步保 证原图与增强图在空间关注区域上的一致性，从而提升模型的域鲁棒性。大量实验结果表明，本 文方法在多个跨域人群计数基准上均取得了优异的泛化性能，验证了所提频域特征解耦与融合策 略的有效性。

Abstract: In cross-domain crowd counting, discrepancies in feature distributions between source and target domains often lead to a significant drop in model generalization perfor- mance. To address this issue, we propose a single-domain generalization method for crowd counting based on frequency-specific feature decoupling. Specifically, we first apply a two-dimensional Fourier Transform to decompose input images into frequency components, explicitly separating high-frequency and low-frequency information. The high-frequency components capture domain-invariant edge and texture details, while the low-frequency components represent the global structural and density distribution information of the image. Considering that density regression involves locally smooth variations that are mainly influenced by low-frequency components, we introduce a Content Error Mask (CEM) to filter domain-specific information from low-frequency features. Furthermore, a high-frequency-guided spatial attention mechanism is de- signed to achieve effective frequency-domain feature fusion. An additional attention consistency constraint is applied to ensure consistent spatial focus between the orig- inal and augmented images, thereby improving cross-domain robustness. Extensive experimental results on multiple benchmark datasets demonstrate that our method achieves superior generalization performance, validating the effectiveness of the pro- posed frequency-specific feature decoupling and fusion strategy.