<p>Pavement cracks present significant segmentation challenges due to their slender morphology, blu-rred edges, and extremely low pixel proportion. Existing methods often rely on fixed-weight fusion strategies, overlooking the scale-dependent signal-to-noise variations in frequency domains and failing to fully leverage directional cues. To address these issues, a crack segmentation network (MHFFNet) is proposed, which performs multi-scale high-low frequency feature fusion based on wavelet decomposition and the Attentive State Space Module (ASSM). First, a Frequency-guided Feature Extraction Module is proposed. By aligning the inherent directionality of Haar wavelet subbands with crack topology, this module guides the optimization of spatial horizontal and vertical convolutions to robustly extract directional structural features. Second, a high-low frequency feature fusion module is proposed to fuse multi-scale features, comprising Low-Frequency Fusion (LFF) and High-Frequency Fusion (HFF). The LFF incorporates ASSM to dynamically integrate multi-scale low-frequency global features, utilizing their semantic information to effectively suppress background noise. The HFF aggregates multi-scale high-frequency local features to capture spatial details and enhance crack edge representation. Third, extensive experiments demonstrate that the proposed model outperforms eight state-of-the-art methods in terms of segmentation accuracy while maintaining real-time inference capability. MHFFNet achieves mIoU scores of 87.93%, 69.75%, and 81.75% on DeepCrack, CrackLS315, and CFD, respectively, and runs at 33 FPS on DeepCrack.</p>