Study on segmentation of fine cracks with water seepage on hybrid towers of wind turbines

To address the engineering challenge of detecting fine cracks on hybrid wind turbine towers, especially against complex water seepage backgrounds, this study aims to explore optimal image segmentation strategies. The core challenges of this task lie in the severe class imbalance caused by the extremely low pixel ratio of crack targets and the visual interference from seepage areas. To this end, a dedicated dataset for this specific scenario, named HTSCD, was first constructed. Subsequently, based on the U-Net segmentation model, this study systematically compared the effects of various combinations of data processing strategies (original, tiled, tiled and augmented) and loss functions (Cross-Entropy, Weighted Cross-Entropy, Dice Loss). Furthermore, to investigate the potential performance improvement from external data, the effectiveness of transfer learning using public crack datasets and programmatically synthesized data was also evaluated. The experimental results demonstrate that the combination of the tiled and augmentated dataset strategy and the Dice Loss function is the optimal solution for this task, achieving the best balance between precision and recall. A key finding is that conventional transfer learning strategies exhibited significant “negative transfer” in this task, where the introduction of external data impaired model performance. This research not only establishes an effective baseline solution for wind tower crack detection in this specific scenario but also provides important practical insights into the limitations of transfer learning for highly specialized visual inspection tasks.