A Novel Edge Detection Method of Blade With Multisupervision for Foreground–Background Confusion Caused by Extreme Illumination

In the production and maintenance of aeroengine blades (AEBs), appearance measurement of AEBs is essential. Vision-based measurement (VBM) using vision sensors can quickly acquire the contour dimensions of AEBs. However, AEBs with smooth curved surfaces are susceptible to lighting conditions, which can reduce the effectiveness and accuracy of VBM. Since edge detection is the basis of VBM methods, the difficulties in VBM of AEBs under the influence of illumination are that edge predictions do not align well with the actual semantic content of the image for some foreground and background confusion situations that occur near the object contour (i.e., overexposed areas exist simultaneously in both the object and background). Therefore, in this article, we propose a method called blade edge detection network with illumination robustness (BEDNet-IR). Our method aims to leverage the information from both salient object detection (SOD) and high-level semantic features to improve the accuracy of edge predictions and robustness to illumination, further enhancing the robustness and accuracy of blade measurement. Additionally, we constructed a challenging dataset which contains many blade images taken under extreme illumination conditions. We conducted experiments using this dataset to evaluate existing edge detection methods alongside our own, demonstrating the feasibility and effectiveness of our approach. Specifically, our method achieved ODS-F scores of 0.797 and 0.815, as well as OIS-F scores of 0.917 and 0.926 across two different dataset partitions. The code and dataset will be released at
https://github.com/VDT-2048/BEDNet-IR
.