Segmentation and constrained quantification method for CFRP surface defects under multi-scale coexisting conditions

Carbon fibre-reinforced polymer composites are widely employed in engineering structures due to their superior mechanical performance. Surface defects can degrade structural integrity and service reliability. However, accurate defect segmentation and reliable quantitative evaluation remain challenging in practical inspection scenarios owing to significant scale variation, severe class imbalance, and low surface-texture contrast. To address these issues, this study investigated the influence of defect-scale distribution on semantic segmentation stability and quantitative evaluation performance. Two datasets were constructed: a multi-scale composite defect dataset and a single-scale defect dataset. Comparative experiments were conducted to analyse the impact of multi-scale coexistence on segmentation performance. A Multi-Scale Class-Aware Segmentation Network was developed by integrating Atrous Spatial Pyramid Pooling, a gated attention mechanism, and a class-aware expert head into a U-Net backbone. In addition, a hybrid loss function combining Focal Cross-Entropy and Focal Tversky was introduced to enhance robustness under class-imbalanced conditions. Furthermore, a segmentation-constrained defect quantification framework was established to refine defect contours and improve the measurement reliability. Experimental results demonstrate that the proposed method enhances segmentation stability under multi-scale coexistence and enables reliable quantitative evaluation of surface defects in composite inspection.