Localisation of defects on carbon fibre surfaces using deep learning

Abstract

This study explores the application of deep learning for localisation and classification of common defects in carbon fibre materials. A Sony IMX250MZR polarisation camera was employed to leverage the polarising properties of CFRP surfaces. However, analysis revealed that the additional polarisation data provided minimal advantages. Instead, promising results were obtained using a standard monochrome output from the sensor. Defects were classified into two categories: “carbon fibre defects” and “foreign bodies”. A dataset comprising of 2400+ annotated instances for each type was analysed using two state-of-the-art deep learning models: YOLOv11-seg,for instance segmentation, and SegFormer for pixel-wise classification. Images were captured under three different illumination conditions including specialised dome lighting and strip lights in ambient environments. Each lighting configuration yielded promising results, with dome lighting demonstrating superior performance. YOLOv11 achieved an average precision score of 0.817 under dome lighting, compared to 0.672 in the least favourable lighting scenario. SegFormer slightly outperformed YOLOv11 in segmentation accuracy, achieving a mean Intersection over Union (mIoU) of 0.742 compared to 0.678 for YOLOv11. The consistently high detection rates demonstrate the potential of both models for reliable identification of critical and minor defects, making them well-suited for industrial quality assurance.