DCE-YOLOv8n: a lightweight attention-enhanced model for weld surface defect detection in low-light scenes

Detecting weld surface defects under low-light conditions presents significant challenges due to limited illumination and subtle defect features. In this study, we propose a Visual Inspection System for Weld Surface Defects (VISFWSD). To support training and evaluation, we construct a dedicated Weld Surface Defect Data Set (WSDDS). Building upon the YOLOv8n architecture, we introduce DCE-YOLOv8n, a lightweight and high-performance model enhanced with Deformable Convolutional Networks (DCNv3) and the Convolutional Block Attention Module (CBAM) to improve multi-scale and fine-grained defect feature extraction. Furthermore, we adopt the Enhanced Intersection over Union (EIoU) loss function to address training sample imbalance and enhance localization accuracy. The proposed DCE-YOLOv8n model achieves superior performance across multiple evaluation metrics compared to baseline models. Specifically, it attains the highest mAP@0.5 of 92.9% and mAP@0.5:0.95 of 64.5%, outperforming YOLOv8n by 2.7 and 3.1 percentage points, respectively. Despite these accuracy gains, DCE-YOLOv8n maintains competitive computational efficiency, with 7.6 GFLOPS and 2.4 million parameters, both slightly lower than those of YOLOv8n (8.0 GFLOPS and 3.1 million parameters). Furthermore, it achieves an inference speed of 200 Frames Per Second (FPS), satisfying real-time processing requirements. These results demonstrate that DCE-YOLOv8n effectively balances detection accuracy, model size, and computational cost, making it well-suited for practical deployment in low-light weld defect inspection tasks. The proposed VISFWSD and DCE-YOLOv8n together provide an effective and efficient solution for weld surface defect detection in low-light environments.