Cross-category few-shot weld defect detection via improved YOLOv8 with stage-wise transfer learning (<italic>invited</italic>)

ObjectiveWeld inspection is essential for ensuring manufacturing quality and structural reliability. However, weld images collected from real industrial environments exhibit large cross-category variations and limited defect samples, making small-sample detection highly challenging. This study aims to develop a high-accuracy and robust weld detection method under few-shot and cross-category conditions based on an improved YOLOv8 framework.MethodsA cross-category small-sample weld detection framework is constructed using YOLOv8 as the baseline. Bilateral filtering and histogram equalization suppress background noise and enhance weld contrast, enabling more discriminative features under limited samples. The Simple Parameter-Free Attention Module (SimAM) is embedded in the backbone to strengthen focus on key weld regions. A staged transfer learning strategy gradually unfreezes backbone layers, mitigating feature distribution discrepancies and improving stability in small-sample training. Ablation and comparative experiments under different dataset scales assess module contributions, robustness, and generalization of the proposed method.Results and DiscussionsThe improved model outperforms the baseline YOLOv8 across all training-sample settings. With 100 training images, it achieves a mAP50 of 0.985 and a mAP50-95 of 0.659, surpassing the baseline by approximately 5.9 and 22.3 percentage points, respectively. When the number of training images is further reduced to 40, the proposed method still maintains high performance, achieving a mAP50 of 0.971 and a mAP50-95 of 0.585, demonstrating strong robustness and generalization ability under limited-sample conditions.ConclusionsThe proposed approach demonstrates superior detection accuracy and robustness under small-sample, cross-category, and complex industrial conditions. Benefiting from its simple structure, enhanced feature representation, and strong transfer adaptability, the method provides reliable support for automated weld inspection systems and exhibits considerable engineering application potential. The framework offers a practical solution for high-precision weld detection with limited data and lays a foundation for extending few-shot learning to broader industrial visual inspection tasks.