Improved YOLOv8n-Based Model for Wind Turbine Blade Defect Detection
TL;DR
First, MS-CBAM is introduced before the SPPF module to enhance channel-wise feature refinement and multi-scale spatial feature extraction and EPConv is replaced with EPConv, which combines efficient multi-scale channel attention with directional pinwheel-shaped convolution to strengthen the representation of weak and elongated defects.
Abstract
Wind turbine blade defect detection in complex environments is challenged by weak defect features, missed crack detections, and false detections caused by background interference. To address these problems, this study proposes an improved YOLOv8n-based detection model. First, MS-CBAM is introduced before the SPPF module to enhance channel-wise feature refinement and multi-scale spatial feature extraction. Second, four stride-2 downsampling convolutional layers in the backbone are replaced with EPConv, which combines efficient multi-scale channel attention with directional pinwheel-shaped convolution to strengthen the representation of weak and elongated defects. Finally, the original CIoU loss is replaced with PIoU v2 to improve bounding-box regression. Experiments on a self-constructed wind turbine blade defect dataset show that the proposed model achieves a precision of 92.1%, a recall of 85.1%, an mAP0.5 of 90.7%, and an mAP0.5:0.95 of 68.0%. Compared with the original YOLOv8n, these values represent improvements of 0.4, 5.9, 3.9, and 5.2 percentage points, respectively. The model contains approximately 3.0 M parameters, requires 8.9 GFLOPs, and achieves a network-forward inference speed of 77.1 FPS on an NVIDIA GeForce RTX 5060 Laptop GPU. Class-wise evaluation further shows that crack AP0.5 increases from 78.4% to 85.0%, while crack AP0.5:0.95 increases from 50.7% to 55.9%. These results demonstrate that the proposed modifications improve the detection and localization of weak and elongated defects while maintaining real-time inference capability on the tested GPU platform.