Advancing AI-Driven Defect Detection in Carbon Fiber Materials | Technical paper
Segmentation Supervised Multi-Head Defect Classification of Virgin and Recycled Carbon Fiber Material in Photometric Stereo Images
| A newly developed segmentation-supervised multi-head neural network is setting new benchmarks in defect classification for virgin and recycled carbon fiber materials. With the increasing adoption of neural networks in digital image-based inspection, researchers face challenges due to carbon fiber’s non-rigid nature and high reflectivity, which complicate robust quality assessment. To address these difficulties, this approach integrates a multi-head, multi-task neural architecture with optimized convolutional layers, enabling shared embeddings for image classification and segmentation in photometric stereo imaging. The framework enhances defect classification by utilizing class activation maps (CAMs) and segmentation outputs as attention mechanisms. Performance evaluations conducted on a publicly available dataset show that this model surpasses standard CNNs and transformer-based methods, achieving over 30% higher accuracy, exceeding previous benchmarks by 4.4%. Key Innovations: – Modular optical sensor: An extended camera sensor with photometric stereo capabilities measure fiber orientation parameters. – Improved classification accuracy: CAM-based attention mechanisms enhance defect classification performance. – Multi-head architecture: A lightweight architecture processes image classification and segmentation concurrently. – Benchmark performance: Extensive evaluations on publicly available datasets validate the model’s superior defect detection. An extended photometric stereo camera setup was developed, capturing multi-modal images that undergo stitching and pre-processing for further analysis. The segmentation architecture applies transposed convolutional layers to extract detailed segmentation features, which are fine-tuned with CAM layers. These segmentations are binarized and used as masks, refining convolution feature extraction for classification tasks. Conclusions and Future Directions: This research introduces a comprehensive framework for defect classification, integrating modular optical sensing, attention-based classification refinements, and a lightweight multi-head architecture. The proposed model improves accuracy while reducing reliance on large datasets, a challenge in industrial applications. Next Steps: The study acknowledges limitations in dataset availability, restricting model generalization across diverse industrial scenarios. Future efforts will focus on image generation using few-shot learning, creating synthetic defect samples to strengthen model scalability. For access to the dataset and additional technical details, visit Zenodo. |
Note: The Zenodo links, this is a dataset that has been created during the MC4 project. All the experiments conducted for the paper used this dataset which is the first and only available dataset in the world for this category. The paper (presented at Sampe Europe 2025) shares the acknowledgements of both the MC4 and COMPASS projects.