A Lightweight Parallel Attention U-Net for Surface Defect Segmentation of Wind Turbine Towers in Visible-Light Images

Wind turbine towers operate in complex outdoor environments, where visible surface anomalies such as cracks, pitting, and honeycombing can develop. Field-acquired visible-light images are commonly affected by illumination variation, shadows, local reflections, surface textures, and structural joints, which makes pixel-level anomaly segmentation difficult. This study proposes a task-oriented lightweight U-Net, termed LPAU-Net, that combines DWConv–PWConv feature extraction, parallel channel–spatial attention with learnable scalar fusion, grouped multi-level feature aggregation, and multi-scale decoding. Defect-free images are included during training, and defective and defect-free samples are evaluated separately to distinguish anomaly segmentation from false-positive suppression. The dataset contains 762 original field images collected from the same nine wind turbine towers at one wind farm during five time-separated acquisition campaigns. Campaigns 1–3 were used for training, Campaign 4 for validation, checkpoint selection, and threshold determination, and Campaign 5 for final evaluation. Thus, Campaign 5 is a later acquisition batch from the same towers and site, rather than unseen-tower or cross-wind-farm validation. Across three independent random seeds, LPAU-Net achieved 89.84 ± 0.08% Precision, 89.24 ± 0.08% Recall, 89.54 ± 0.08% F1-score, and 81.07 ± 0.13% Defect IoU on defective Campaign 5 images, with 4.34 M parameters, 14.8 G FLOPs, and 32.81 FPS under the reported desktop-GPU benchmark. On the 30 defect-free Campaign 5 images, the average false-positive area ratio was 0.42 ± 0.03%, and the image-level false-alarm rate was 10.00 ± 3.33%. The results indicate a balanced accuracy–complexity trade-off within the evaluated cross-time-campaign setting. However, strong light, low light, shadows, and reflections were not evaluated as independent subsets, so condition-specific robustness improvement cannot be quantified. Because all visible anomalies were merged into one binary defect class, the model localizes anomalous regions but does not classify cracks, pitting, honeycombing, or other defect types.