High–low frequency feature fusion network for pavement crack segmentation in complex environments

For routine road maintenance, accurate and efficient crack detection holds significant research value and practical importance. With advances in computer vision, image processing-based crack detection has achieved promising results. However, in complex environments, factors such as illumination and weather introduce noise interference, especially in high-resolution crack images where fine fracture features show heightened complexity and diversity. This often leads to degraded detection accuracy in such challenging scenarios. To address this, we propose FE-SegNeXt, a convolutional neural network segmentation framework tailored for high-resolution crack detection in complex environments. The algorithm leverages distinct characteristics of high- and low-frequency features in crack images: average pooling separates these features, while a dedicated Frequency Collaborative Enhancement Module (FCEM) independently enhances high- and low-frequency components before fusing multi-band information. This design significantly improves the model’s ability to extract subtle cracks from high-resolution images under noisy conditions. Additionally, we introduce a Locally Enhanced Feed-Forward Network (LE-FFN) to amplify perception of weak crack signals in local regions, further refining fine-grained feature extraction. Experimental results on the public datasets Sun520, Rain365, and BJN260 demonstrate that the proposed method achieves F1-scores of 61.55%, 62.57%, and 60.31%, respectively, outperforming existing crack detection algorithms. • FE-SegNeXt detects fine cracks in high-res road images via frequency-aware feature fusion. • A dual-branch design captures edges and global context from high- and low- frequency features. • LFFN expands receptive fields and enhances detail for subtle and broken crack detection.