Segmentation Aware Attention Mechanism for Defect Classification of both Virgin and Recycled Carbon Fiber Fabric | from COMPOSITE 2025
Using neural networks for photometric stereo based surface inspection is fast emerging, especially for carbon fiber matrix (before binding with resins) / fabric.
However, the non-rigid nature and high reflectivity of the carbon fiber material makes them very difficult to perfrom robust digital image based quality inspection.
To overcome this issue:
1) a dataset was collected with photometric stereo feature modalities and open sourced
2) studied the classification performance and the dataset
3) propose a segmentation supervised multi-head defect lassification model.
Several standard vision classifier models are retrainded on the created dataset to benchmark the permormance of our model.
This paper demonstrates that a segmentation supervised multi-head neural network outperforms the benchmark.
The realized classifier model achieves over 30% higher accuracy compared to standard CNNs and transformer models. The dataset for classification and its parser its avalaible at this link.
