SFCMs-DETR: An Enhanced DEtection TRansformer for Small-Scale Defect Detection in Fibre Composite Materials
Abstract
Fibre composite materials (FCMs) are widely used in modern engineering, yet they remain susceptible to critical structural defects throughout their service lifecycle. Owing to its end-to-end detection paradigm, the DEtection TRansformer (DETR) has attracted increasing attention for FCMs defect detection. However, the inherent low-resolution feature representations extracted by the standard DETR limit its ability to identify small-scale defects with subtle morphological characteristics for FCMs. To overcome this limitation, a small-scale FCMs defect detection DETR (SFCMs-DETR) framework is proposed in this paper. The proposed SFCMs-DETR framework incorporates a bidirectional feature pyramid network to enhance multi-scale feature fusion. In addition, a brightness intensity-based penalty term is introduced into the loss function to capture the characteristic intensity patterns of FCMs defects and enhance the model’s attention to defect regions. Experiments are carried out on a public FCMs defect dataset for small-scale defect detection with complex backgrounds. Experimental results demonstrate the superiority of the SFCMs-DETR framework over representative FCMs defect detection methods, including the standard DETR, YOLOv8, YOLOv12, and the Faster R-CNN. The proposed SFCMs-DETR framework exhibits substantial potential for large-scale industrial and commercial deployment of FCMs defect detection in the future.