A lightweight real-time detection transformer model for surface defect detection systems

Abstract

Accurate and efficient detection of surface defects is essential for production and infrastructure monitoring. A lightweight real-time surface defect detector is proposed, the Surface Defect Detection Real-Time DEtection TRansformer (SDD-RTDETR). It builds upon the Real-Time DEtection TRansformer v2 (RT-DETRv2). Key innovations include Re-Parameterized Partial Convolution (RPConv) within the BasicBlock_RPConv, which minimizes computational workload and memory requirements while boosting performance. The model also proposes the Efficient Multi-Scale Attention-based Feature Interaction (EMSAFI) module to strengthen feature extraction capabilities and employs the lightweight fusion architecture LiteScaleNeck to optimize feature fusion. Additionally, the Inner-Minimum Point Distance Intersection over Union (Inner-MPDIoU) loss refines bounding box regression, further improving model performance. The experimental findings reveal that SDD-RTDETR excels across multiple surface defect datasets. In contrast to the benchmark model, this approach improves detection accuracy while decreasing parameters by 34.6 % and computational complexity by 23.0 %, validating its adaptability and generalization ability in surface defect testing. With its lightweight structure and superior capability, SDD-RTDETR provides an effective approach for large-scale, immediate inspection, driving automation in quality control and infrastructure monitoring.