An efficient hybrid weld seam detection network based on knowledge distillation and model pruning

Abstract Accurate weld type recognition and stable keypoint localization are essential for automated weld seam inspection, yet real-time deployment is still challenged by the high computational cost of dense prediction models and visual disturbances such as arc glare and spatter. This paper presents LDCNet, a lightweight compression framework for DenseNet-CenterNet. The framework considers both the structural dependencies of dense connections and the spatial sensitivity of weld keypoint localization during model compression. In the pruning stage, DepGraph-based structured pruning is applied to achieve dependency-consistent channel compression in the densely connected network while preserving the functional integrity of the classification, heatmap, and offset prediction branches. In the distillation stage, multi-level knowledge distillation is used to align intermediate features together with heatmap and offset responses, thereby recovering the representation and localization performance degraded by pruning. Experimental results show that LDCNet maintains good weld type recognition and keypoint localization performance while substantially reducing the model size. Experiments on a dataset of 2,370 weld images covering fillet, half-V groove, and V-groove welds show that LDCNet reduces the number of parameters by 69% and reduces the inference latency by 64.1%, while achieving 96.4% classification accuracy. These results show that LDCNet achieves a good balance between accuracy and computational efficiency for weld seam detection.