.png){kind=icon}
.png){kind=icon}
Return
Multi-scale class activation map for weakly supervised defect segmentation
.png){kind=icon}
PRE
AI
DOI：10.1007/s10044-026-01660-9.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Surface defect detection is vital to industrial quality assurance. Deep learning excels at inspection but relies on costly pixel-level or bounding-box annotations, constraining dataset scale. Weakly supervised learning reduces labeling cost yet localizes defects less precisely than full supervision. We propose MSFF-Net, a multi-scale feature-fusion weakly supervised localization network with two parts: a classification model and Probes-CAM. The classifier uses a multi-branch design with transfer learning and augmentation. A primary branch employs a pretrained backbone to capture context, while auxiliary branches expand capacity. Training is phased: once the primary branch is trained, its parameters are frozen; auxiliary branches are optimized sequentially to learn mid-level details and enable multi-level fusion. Probes-CAM refines activation mapping by exploiting features from both primary and auxiliary branches. After bilinear interpolation and weighted fusion, an adaptive threshold yields pixel-level masks with improved localization fidelity. Evaluations on DAGM and Magnetic-Tile-Defect show Probes-CAM raises IoU by 4.97% over the baseline, improving pseudo-label quality and segmentation precision. By combining transfer-learned context, auxiliary-branch detail, and CAM refinement, MSFF-Net narrows the localization gap between weak and full supervision while preserving labeling efficiency crucial for scalable industrial deployment. Code and models are extensible and generalize across datasets and defect types. The code is available at https://github.com/Jyb0592-code/Torch_probe_cam.git
Keywords:
Industrial defect detection
Weakly supervised learning
Semantic segmentation
Explainable AI
Class activation maps
Journal
IF:
2
Papers:
1.9K
Citations:
1.9K
View more organizations