LSDF-Net: An efficient lightweight defect detection method for ultrasonic welding surfaces

This paper proposes LSDF-Net, a lightweight and high-speed detection network designed to address the challenges of insufficient accuracy and high computational cost in ultrasonic welding surface defect detection. Built upon the YOLOv8 architecture, LSDF-Net integrates a Dynamic Surface Detail Fusion Module (DSDFM) to enhance multi-scale feature representation and introduces a Lightweight Shared Convolution and Separate Batch Normalization detection head (LSCSBD) to reduce parameters and accelerate inference. In addition, a LAMP-based pruning strategy is applied, which achieves a 67% reduction in model size and a 48% reduction in computational cost with almost no performance degradation. Experimental results on both a self-constructed ultrasonic welding defect dataset and the public NEU-DET dataset demonstrate that LSDF-Net achieves the best overall performance, striking an excellent balance between accuracy and real-time inference. These results highlight its strong potential for real-time industrial defect detection applications.