Pixel U-Net: an improved version of U-Net for binary segmentation of wind turbine blades

Wind turbine technicians face significant risks of fatal injuries, which can be mitigated by utilizing drones to capture images of wind turbine blades (WTBs) for remote inspection and maintenance. Different computer vision and image processing techniques can be applied to these captured drone images to automate the process of WTB inspection. However, the captured drone images consist of challenging backgrounds, due to which the WTB area needs to be extracted from the image as a pre-processing step. This paper introduces Pixel U-Net, an enhanced U-Net architecture tailored for binary segmentation of WTB images, improving segmentation accuracy with pixel shuffle and unshuffle operations. Evaluated against baseline U-Net architecture and its variations using the publicly available Blade30 dataset, Pixel U-Net achieves an average validation accuracy of 99.0% and surpasses existing methods in WTB image segmentation. Additionally, novel architectural variations, namely U-Net + HS-Block + Pixel Unshuffle and Pixel U-Net + Attention, have also been proposed in this study, which exhibit superior performance with an average training loss of 0.012 and an average testing accuracy of 98.3%, respectively. Qualitative comparisons of the results further highlight the efficacy of deep learning-based segmentation techniques in advancing wind turbine inspection and maintenance practices.