DFI-YOLOv8 based defect detection method for fan blades

To achieve real-time and accurate detection of blade defects in wind turbines, a wind turbine blade defect detection algorithm based on the YOLOv8 network model is proposed DFI-YOLOv8. Firstly, in response to the fixed shape and size of convolution kernels in ordinary convolutions, and the defects of complex background, high noise, and strong interference on the surface of wind turbine blades, an expandable deformable convolution network DCN was added to the Backbone network, and a DCN_C2f module was constructed to enhance the feature extraction ability of the model under complex background conditions. At the same time, an EMA attention mechanism was introduced to highlight important feature information on the surface of wind turbine blades; Secondly, to address the issues of traditional detection models being bulky and difficult to deploy, a lightweight module called FasterNet is embedded in the Neck layer to reduce redundant computation and memory access, while more effectively extracting spatial features; Finally, to address the issue of varying sizes of defects in wind turbine blades during the detection process, which can lead to missed detections due to smaller defects, a hollow Inception SimAM convolution module is proposed to effectively improve the model's ability to recognize small targets. The experimental results show that the accuracy of DFI-YOLOv8 can reach 80.8%. Compared with the mainstream YOLO series algorithms, mAP50 has significantly improved, while the number of parameters has also been reduced to a certain extent. This fully demonstrates the superior performance and efficient accuracy of DFI-YOLOv8 in detecting defects in wind turbine blades.