Visual Detection of Surface Defects in Wind Turbine Blades Using Two-Step Image Segmentation

Resumo

Automated visual inspection of wind turbine blades (WTBs) using drone imagery is a crucial yet complex task, primarily due to the large scale of the structures, varying lighting conditions, and the wide variety of potential surface defects. These challenges demand solutions that are both precise and scalable. To address this, we propose a robust two-step segmentation pipeline to efficiently localize and analyze defects in real-world drone images of WTBs. Our approach leverages two cascaded deep learning models operating at different spatial resolutions. In the first stage, a global segmentation model identifies and isolates the blade from its background, terrain, or turbine tower by generating a binary mask, significantly reducing the search space, focusing the analysis on the most relevant region, and minimizing computational overhead. In the second stage, the extracted blade area is subdivided into overlapping high-resolution patches that retain detailed texture and structural information necessary for accurate defect recognition. A segmentation network (U-Net) is then applied to these patches, producing multi-channel binary masks where each channel represents a specific defect class, such as adhered material, erosion and mechanical impact. This hierarchical architecture offers several key advantages: enhanced computational efficiency by concentrating fine-grained analysis on smaller regions of interest; modularity that allows each model to be optimized independently; and support for multi-class segmentation, enabling simultaneous detection of multiple defect types—an essential feature in real-world inspections, where more than one damage is typical. Moreover, the spatially coherent masks produced by both stages allow for geometric post-processing, making it possible to reconstruct the location of defects across the entire blade and enabling the measurement of geometric attributes such as defect area, relative position, and clustering. Such spatial insights are vital for guiding maintenance operations and supporting predictive maintenance strategies. Our pipeline balances accuracy and efficiency by decomposing the task into the global blade and local defect segmentation. The proposed method offers a scalable and versatile framework for reliable WTB inspection, with strong potential for real-world deployment and further research exploration.