CM-YOLO: an improved YOLOv8 for defect detection and classification of CFRP composites using infrared thermography

Abstract

Driven by the pursuit of higher operating efficiency and better structural safety, carbon-fibre-reinforced polymer (CFRP) composites are widely adopted in critical aerospace structural components. Infrared thermography has emerged as a powerful non-destructive technique for subsurface defect inspection in composite materials.Facing blurry defects and noisy raw infrared thermograms in automate defect recognition, this work presents an enhanced YOLO-based network, named CM-YOLOv8n, which integrates a Channel Refinement Attention (CRA) module and a Multi-Dimensional Feature Fusion (MDFF) module for high-precision infrared defect classification. Long-pulse thermography experiments were carried out on CFRP plates embedded with irregular artificial defects, and a dedicated dataset was constructed via data augmentation and pixel-wise labelling of defective regions. The refined CM-YOLOv8n model was trained on the preprocessed dataset to extract discriminative defect features. Quantitative experimental results demonstrate excellent performance of the CM-YOLOv8n model with an mAP@0.5 of 89.77%, a Precision of 97.73% and a Recall of 80.97% for classification of defects. Ablation experiments confirm that the embedded CRA and MDFF modules jointly boost model performance by suppressing thermal noise and sharpening defect edges in infrared frames. Furthermore, comparative experiments against state-of-the-art lightweight YOLO architectures reveal that our model achieves superior classification accuracy while retaining a compact parameter footprint. These merits render CM-YOLOv8n suitable for automated quantitative assessment in industrial non-destructive testing workflows.