Paper Details
Abstract
This study presents a customized pipeline that combines image preprocessing and seg- mentation to enhance the accuracy of surface defect localization on wind turbine blades. Aerial imagery captured by UAVs is often degraded by non-uniform lighting, visual noise, and poor contrast, which impairs the reliability of automated inspection systems. To mitigate these issues, we introduce a domain-adapted preprocessing strategy that im- proves structural definition while filtering out irrelevant background features. The refined images are subsequently analyzed using the Segment Anything Model (SAM), a prompt- driven segmentation framework known for its versatility across diverse visual domains. In this work, SAM is tailored to the wind turbine inspection context through the use of targeted prompts and optimized parameters that reflect characteristic defect signatures. Evaluation is conducted on a manually labeled dataset comprising real-world examples of turbine blade damage under various environmental conditions. The experimental re- sults demonstrate that the integration of preprocessing steps markedly enhances the seg- mentation accuracy of SAM, enabling precise identification of defects such as cracking, erosion, and material loss. The proposed solution improves both segmentation fidelity and model resilience, offering a scalable approach for early-stage fault detection in wind energy systems.