EfficientDet-based surface defect detection algorithm for large rotor room ring parts

In view of the low detection accuracy and large model parameters of current general object detection algorithms for surface defects in the annular components of wind turbine rotor housings, this paper proposes an improved surface defect detection algorithm for rotor housings based on EfficientDet. The main improvements include replacing the Swish activation function and BN layer in the MBConv module with Mish activation function and GN layer; adding dilated convolution modules to enhance the feature extraction capabilities of the backbone network; redesigning the BiFPN network by increasing cross-level and top-down information connections to enhance the feature fusion capabilities of the BiFPN network, and improve the ability to identify small defects on the surface of rotor housing components. Experimental data shows that the improved EfficientDet network model achieves a mean average precision (mAP) value of 77.69% on the NEU-DET dataset, an improvement of 4.83% compared to the original network model, with the smallest model size, meeting real-time detection requirements of steel surface defects in practical applications.