Wavelet Decoupling and Fusion Network for Aero-Engine Blade Defect Detection

Abstract

Turbine blades in aerospace engines are critical components directly related to flight safety, where even minor surface defects can cause severe failures. However, such defects usually present low contrast, irregular geometry, and large-scale variations, making existing segmentation methods prone to missed detections and inaccurate boundary localization. To address these issues, this article proposes a wavelet decoupling and fusion network (WDFNet) for high-precision turbine blade defect segmentation. A region of interest attention enhancement (RoIAE) module is first introduced to adaptively emphasize defect-related features while suppressing background interference. Then, a wavelet frequency decoupling (WFD) block is incorporated into the decoder to explicitly separate structural context and fine-grained edge details, enabling accurate reconstruction of multiscale defect boundaries through adaptive frequency fusion. Experiments on the aero-engine blade defect detection (ABDD) dataset show that WDFNet achieves a mean intersection over union (mIoU) of 95.73 %, with a precision of 96.23 % and a recall of 98.11 %, outperforming advanced posthoc methods while maintaining high computational efficiency. Furthermore, WDFNet is deployed in a fully automated defect detection system supporting autonomous data acquisition and real-time inference. The project source code, dataset examples, and a demonstration video of the automated defect detection system are available at https://github.com/boningboning/defect-detection-demo.git