MDFANet: Multidomain Feature Alignment Surface Defect Detection for Aircraft Engine Turbine Blades

Aeroengines are critical components of aircraft, and defect detection during the manufacturing stage entails a substantial workload. Currently, fully supervised learning methods necessitate extensive datasets to achieve high-precision detection. The time-consuming nature and complexity of data collection, however, pose significant challenges in constructing extensive datasets. In order to deal with this concern, we proposed a multidomain feature alignment network (MDFANet). First, we constructed a source domain dataset, which comprises six datasets of varying sizes, encompassing defect categories found on aircraft engine turbine blade surfaces and incorporating additional defect types. Then, we introduced the domain adaptation module into the YOLOv7 framework, which actively adjusts the feature distribution between domains during training, thereby reducing the impact of domain shift. This enhancement ensures the model’s generalization ability across different defect scenarios and provides potential recognition capability for future defects that may appear on the blade surface. Additionally, we designed a novel feature fusion module in the backbone layer, named “multiscale feature fusion” (MFF) module, which facilitates the extraction of image features from different domains, thereby capturing fine-grained details and contextual information more comprehensively. Additionally, a small defect detection head was integrated into the module head to further improve the detection capability for micro defects on aircraft engine turbine blade surfaces. Finally, we introduced the MPDIoU loss function into YOLOv7, enabling faster convergence for training and improving the robustness. Extensive experimental evaluations demonstrate that, while balancing inference efficiency and detection accuracy, our approach achieves mean Average Precision (mAP) of 0.469 across all categories and mAP50 of 0.863, significantly enhancing both the detection speed and accuracy of aircraft engine turbine blade surface defects.