Context-Guided Hard-Negative Background Suppression for Crack Segmentation on Complex-Texture Farmland Roads

Field roads in high-standard farmland are continuously exposed to agricultural machinery loads, environmental degradation, and limited maintenance conditions. Consequently, pavement cracks in these environments are often characterized by slender morphology, low contrast, irregular boundaries, and highly complex background textures. Although existing U-Net-based crack segmentation models enable pixel-level detection, they are prone to misclassifying crack-like background patterns as genuine cracks in scenes containing coarse aggregates, tire marks, shadows, repair edges, and strong pavement noise, thereby increasing false-positive predictions. To address this issue, this study proposes a Context-Guided Hard-Negative Background Suppression Network (CGHN-Net). Built upon a Squeeze-and-Excitation Channel Attention-U-Net (SE-U-Net)backbone, CGHN-Net introduces a Context-Guided Skip Gate (CGSG) between the encoder and decoder. The CGSG exploits high-level semantic information from the decoding stage to spatially filter low-level skip features, thereby suppressing interference from complex pavement textures during feature fusion. In addition, a Hard-Negative Background Suppression Loss (HNBS Loss) is designed to select hard negative samples with relatively high predicted crack probabilities and large losses from ground-truth background regions and impose additional constraints on these samples, further reducing high-confidence spurious crack responses. The experimental dataset was constructed from two-dimensional grayscale pavement images acquired by a multifunctional pavement inspection vehicle and contained 2,235 samples, including 1,684 training images, 464 validation images, and 87 test images. Compared with representative semantic segmentation models, including U-Net, Attention U-Net, UNet++, and DeepLabV3+, CGHN-Net achieved a Dice coefficient of 0.8682 ± 0.0055, an Intersection over Union (IoU) of 0.7791 ± 0.0102, a precision of 0.8780 ± 0.0161, and a recall of 0.8734 ± 0.0266 over three random-seed trials, while reducing the mean false-positive area ratio (FP Area Ratio) to 0.0042 ± 0.0008. Ablation experiments demonstrate that CGSG effectively improves the discrimination of crack-related regions, whereas HNBS Loss further suppresses spurious crack predictions in complex-textured backgrounds when used jointly with CGSG.