L-RTDETR: A Lightweight Real-Time Object Detection Algorithm for Defect Detection

In the current deep learning research landscape, image processing technology faces significant challenges in lightweight defect detection despite its broad application in object detection. This study introduces the Lightweight Real-Time DEtection TRansformer (L-RTDETR) model, achieving a balance between high accuracy and model lightness. The model employs a Large Selective Kernel Network (LSKNet) based lightweight backbone, reducing network size and parameters. It also integrates Deformable Attention (DAttention) based Intra-scale Feature Interaction to enhance detection speed with a sparse attention mechanism. By combining L1, Generalized Intersection over Union (GIoU), and Normalized Wasserstein Distance (NWD) loss functions with weighted metrics, the model’s adaptability for object detection tasks improves. Experimental results on the Northeastern University Defect Detection dataset show that L-RTDETR achieves a 35.8% reduction in model size and increases detection speed to 61.6fps, outperforming the conventional RTDETR algorithm. It also shows a 2.5% improvement in mean Average Precision at 0.5 (mAP@0.5) and a 1.7% increase in mAP@0.5-0.95. These results underscore the model’s efficiency in lightweight defect detection.