Surface Defect Recognition of Wind Turbine Blades Based on Improved YOLOX-X Model

Abstract

In order to solve the problem of small data sets and small detected targets in image detection of wind turbine blades. In this paper, we propose an improved YOLOX-X model. Firstly, we use a variety of data set enhancement methods to solve the problem of small data sets. Secondly, an improved Mixup image enhancement method is proposed to enrich the image background. Then, the attention mechanisms of ECAnet and CBAM are introduced to improve the attention of important features. Furthermore, the IOU_LOSS loss function in the original model is replaced with CIOU_LOSS in this paper to improve the positioning accuracy of small target. Last but not least, the overall network uses the Adam optimizer to accelerate network training and recognition. The effectiveness of algorithm is evaluated on a data sets captured by a UAV in a wind farm. Compared with the original YOLOX-X model, our algorithm improves mAP by 4.55%. In addition, compared with other types of YOLO series networks, it is proved that our model is superior to other algorithms.