Wind Turbines Surface Damage Automatic Detection Using YOLOv8 with Specialized Backbone Modification

This work is devoted to incorporating specialized self-attention blocks into deep neural network-based models for detecting and quantifying damage across various components of wind turbines using images captured by unmanned aerial vehicle cameras. In our study, we introduce YOLOv8 backbone modification using a specialized self-attention mechanism, tailored to the specific characteristics of the input data. These modifications aim to enhance the model’s ability to accurately identify and assess the extent of damage from the complex visual data provided by drone imaging. To demonstrate the effectiveness of our proposed solution, we also publish an annotated dataset that we have compiled, which includes images of wind turbines captured by drone cameras. This dataset serves as a valuable resource for training and testing our models. Our solution, evaluated on test subsets of our dataset, has shown state-of-the-art results (mAP50-95 = 0.83234), surpassing most of widely used methods in performance metrics.