Wind Turbine Actual Defects Detection Based on Visible and Infrared Image Fusion

It is of paramount importance to conduct accurate inspections of wind turbine blades to identify and address any defects. However, traditional visual inspection methods are often lacking in intelligence, have high rates of false detections, and are relatively inefficient. Conventional image-based detection methods are also not capable of distinguishing between actual defects, such as coatings falling off, and false positives, such as dust, urine, or feces, in the abnormal areas of images. In this study, we propose a novel approach for distinguishing actual defects in wind turbine blades through the implementation of an efficient data-processing method and a feature fusion module for identifying potential actual defects. By utilizing multiple forms of feature fusion, we are able to effectively eliminate incorrectly characterized defects. Our proposed "Regression Crop" data-processing method enables the automatic selection and cropping of relevant areas of wind turbine blades in the original images, while our adaptive feature fusion module for RGB and infrared images improves classification and localization accuracy for actual defects. Experimental results indicate that our approach achieves optimal results, with the "Regression Crop" data-processing method resulting in a significant improvement in detection accuracy and the adaptive feature fusion module increasing the precision of actual defects to 99%. Furthermore, the adaptive feature fusion module is easily integrated into advanced object detectors such as YOLOv7 to improve their accuracy.