Ultrasound welding technology is widely applied in the field of industrial manufacturing. In complex working conditions, various factors such as welding parameters, equipment conditions and operational techniques contribute to the formation of diverse and unpredictable line defects during the welding process. These defects exhibit characteristics such as varied shapes, random positions, and diverse types. Consequently, traditional defect surface detection methods face challenges in achieving efficient and accurate non-destructive testing. To achieve real-time detection of ultrasound welding defects efficiently, we have developed a lightweight network called the Lightweight Attention Detection Network (LAD-Net) based on an attention mechanism. Firstly, this work proposes a Deformable Convolution Feature Extraction Module (DCFE-Module) aimed at addressing the challenge of extracting features from welding defects characterized by variable shapes, random positions, and complex defect types. Additionally, to prevent the loss of critical defect features and enhance the network's capability for feature extraction and integration, this study designs a Lightweight Step Attention Mechanism Module (LSAM-Module) based on the proposed Step Attention Mechanism Convolution (SAM-Conv). Finally, by integrating the Efficient Multi-scale Attention (EMA) module and the Explicit Visual Center (EVC) module into the network, we address the issue of imbalance between global and local information processing, and promote the integration of key defect features. Qualitative and quantitative experimental results conducted on both ultrasound welding defect data and the publicly available NEU-DET dataset demonstrate that the proposed LAD-Net method achieves high performance. On our custom dataset, the F1 score and mAP@0.5 reached 0.954 and 94.2%, respectively. Furthermore, the method exhibits superior detection performance on the public dataset. © 2024 Elsevier B.V.