Data-centric deep learning for defect detection in fiber composite materials: A review of algorithms, challenges, and emerging trends

Abstract

Fiber composite materials (FCMs) have been widely used in high-performance engineering systems owing to their high strength-to-weight ratio, corrosion resistance, and excellent mechanical properties. Reliable defect detection is therefore essential for ensuring structural integrity, operational safety, and service-life extension. Deep learning (DL) has recently shown strong potential for automated FCMs defect detection by enabling powerful feature extraction, accurate localization, and robust representation learning from image data. Nevertheless, its performance remains highly dependent on sufficient, high-quality, and well-annotated training data, which are often difficult to obtain in practical detection scenarios. This paper presents a comprehensive review of image-based DL approaches for FCMs defect detection under different data conditions. Representative R-CNN-based, YOLO-based, and Transformer-based models are first reviewed for sufficient and high-quality data settings. DL-based denoising, data augmentation, and transfer learning strategies are then analyzed for limited or low-quality data settings, with emphasis on Autoencoders, Generative Adversarial Networks, Variational Autoencoders, Diffusion Models, and pretrained model adaptation. Key challenges and future directions are further discussed, including long-term structural health monitoring, physics-guided defect detection, explainable artificial intelligence, multimodal data fusion, and large language model-assisted detection. By organizing existing studies from a data-centric perspective, this review provides a structured reference for developing reliable, interpretable, and deployable DL-based defect detection systems for FCMs.