Mengwu Wu

Mengwu Wu

Auto-generated profile, based on publicly available information

Mengwu Wu

School of Automotive Engineering and the State Key Laboratory of Light Superalloys, Wuhan University of Technology, Wuhan, China

Research Focus & Metrics

Academic Career Length

1Year (2026-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
26 May 2026 in IEEE Transactions on Instrumentation and Measurement
Verified
Journal Article
Closed access
QGQi Gao
Qi Gao
MWMengwu Wu
Mengwu Wu
Lin HuaLin Hua
Lin Hua
XQXunpeng Qin
Xunpeng Qin
In the realm of surface defect detection for automotive forgings, the identification of surface cracks holds paramount importance, as it is crucial for ensuring both the surface integrity and internal quality of the components. Once defects are detected, the affected areas of automotive forgings must undergo grinding and polishing, a process that relies on accurately determining 3D spatial location of these defects. Consequently, the 3D reconstruction of surface defects on forgings emerges as an indispensable step. However, the intricate morphology of surface cracks in forged components presents a significant challenge for automated detection. To address this issue, this study proposes an innovative cascade framework comprising segmentation-detection-segmentation. The first segmentation module primarily segments the forged component, the detection module aims to localize areas of surface defects in the forging, while the second segmentation module precisely extracts pixel-level features within the crack region. Building upon the accurate segmentation results, the system subsequently performs stereo matching and 3D reconstruction processes, ultimately yielding precise 3D coordinates of the crack defects. The experimental results demonstrate that the U-Net model achieved a mIoU score of 99.03%, providing strong quantitative evidence of its effectiveness in generating precise and reliable segmentation results for the forging components. The YOLO11 object detection model achieved a mAP@0.5:0.95 of 85.9%, outperforming YOLOv5s (82.0%) and YOLOv8s (83.1%). The model achieves an impressive accuracy of 99.3%, with a mAP@0.5:0.95 reaching 85.9%. Furthermore, the U2-Net segmentation model achieved a Dice coefficient of 95.31%, the highest among the three evaluated segmentation models, attaining a Dice similarity coefficient of 95.31%. Quantitative evaluation of the reconstructed cracks reveals high precision, with relative deviations of less than 4% from ground truth for both length and width measurements.