YOLOv8-LC9: a modified YOLOv8 for weld surface detection

Abstract In order to address the computational resource constraints experienced in weld inspection equipment, this study proposes YOLOv8-LC9, a lightweight real-time detection model. The model is built upon ResNet, integrating depthwise, group, and pointwise convolutions to construct lightweight feature extraction modules (C9 and C3Ghost). The C9 module replaces the backbone network of YOLOv8, while the C3Ghost module substitutes the C2f module in the neck network. In this configuration, standard convolutions are replaced by depthwise convolutions. This design enhancement was shown to improve model performance while reducing complexity. Experiments demonstrated that YOLOv8-LC9 outperforms the baseline YOLOv8 in terms of parameters, FLOPs, and model size across n/s/m/l/x variants. The x-model demonstrated the most significant improvements, with a reduction in parameters to 27.361%, FLOPs to 29.29%, and model size to 27.65% of the original, while achieving gains of 2.53% (precision), 1.181% (recall), 1.682% ([email protected]), and 2.669% ([email protected]:0.95). The time taken for single-image detection was shown to decrease from 10.1 ms to 5.2 ms. The model achieves a balance between high accuracy and accelerated inference, thus enabling efficient real-time weld seam inspection.