Real-Time Detection of Subsurface Defects in CFRP Using Line-Laser Scanning Thermography and CSDNet

Carbon fiber reinforced polymers (CFRP) are widely used in aerospace and other fields but are susceptible to defects during manufacturing and service. Infrared thermography-based non-destructive testing enables non-contact, large-area inspection; however, most existing approaches rely on offline processing, limiting their applicability to online scenarios. To address this limitation, a real-time subsurface defect detection method for CFRP based on line-laser scanning thermography is proposed. First, real-time differential preprocessing is applied to suppress the thermal trailing effect. Then, a lightweight detection network, CSDNet (CFRP Subsurface Defect Detection Network), is developed, consisting of three core components: the C2f-RepViT-EMA module, the Lightweight Multi-branch Modulated Feature Pyramid Network (LM²FPN), and the Lightweight Quality-aware Detail Head (LQDH). These components reduce noise interference in infrared images, enhance multi-scale feature representation, and suppress low-quality detection boxes, respectively. Experiments on the T300 CFRP dataset demonstrate that the proposed method achieves a mAP50 of 0.984 with 1.4 million parameters, 5.0G FLOPs, and a deployment inference speed of 30 FPS. Compared with various lightweight detection models, the proposed method achieves a better trade-off between accuracy and efficiency, enabling synchronized scanning and detection and providing effective support for real-time online inspection of CFRP subsurface defects.