Ultrasonic welding is widely used in high-precision manufacturing because of its environmental benefits and high efficiency. Ultrasonic welding surface defects have significant large span and a background noise. Currently, the accuracy and computational complexity of deep learning-based industrial surface defect detection networks are challenging to meet the requirements. To address these problems, this research proposes a lightweight ultrasonic welding surface defect detection network (LUWDNet) for surface defects in ultrasonic welding wire harness terminals. Firstly, an efficient multi-scale residual block (EMR) is proposed, which uses residual structure and reparameterization technology to improve the feature extraction capability and enhance the network’s ability to extract multi-scale features. Subsequently, a deformable directional residual block (DDR) is proposed, which uses a bidirectional convolution kernel to expand the model’s receptive field. Reduces the impact of noise generated by background information in the model. Moreover, a hierarchical cascade feature fusion block (HCFF) is proposed, which uses dynamic channel adjustment to reduce network computation so that network computation and model size can be better adapted to edge deployment. Finally, the Robust Feature Downsampling Module (RFD) is applied in the network to mitigate the problem of small target loss and excessive noise generated by small targets during the downsampling process. To validate the effectiveness and generalizability of LUWDNet, it achieved 94.2% precision and 88.6% mean average precision (mAP50) with 2.6M parameters and a 6.5M model size in the surface defect dataset of ultrasonic welded wire harness terminals (UWSD). In addition, the satisfactory performance on the public dataset NEU-DET demonstrates the generalizability and stability of LUWDNet, showing its broad prospects in the field of industrial defect detection. |
