WELD-DETR: A Real-Time Welding Defect Detection Framework with Multi-Scale Feature Fusion and Multi-Kernel Perception Optimization

Welding technology plays a crucial role in manufacturing, aerospace, construction, and military industries, where the quality of welds directly impacts the safety and reliability of overall structures. Therefore, developing a real-time welding defect detection framework based on deep learning is of paramount importance. However, existing detection methods suffer from limitations such as insufficient real-time performance, high miss rates for small defects, and poor adaptability to complex working conditions. To address these challenges, this paper proposes a welding defect detection framework named WELD-DETR, which incorporates multi-scale feature fusion and multi-kernel perception collaborative optimization. First, we introduce a novel hierarchical feature pyramid (HFPS) structure that effectively combines low-level high-resolution features with high-level semantic features, significantly improving the detection rate of micron-level cracks and pores. Secondly, we innovatively design a multi-kernel perception wavelet convolution (MPWC) module to enhance the model's ability to respond to edge features and fine textures at various scales. Finally, to further boost the model's generalization capability, we construct an industrial-grade welding dataset encompassing five common defect types and propose a cross-condition training strategy based on transfer learning. Experimental results show that WELD-DETR achieves an mAP@0.5-0.95 of 98.2%, a precision of 96.8%, and an inference speed of 58 FPS on an RTX 2060 GPU. Moreover, it exhibits superior detection accuracy and real-time performance in complex industrial scenarios such as high noise and strong reflections, outperforming existing state-of-the-art methods in accuracy. These results underscore WELD-DETR's potential to support intelligent welding quality assurance and process optimization in real-world applications.