Research on HCSE-YOLO Algorithm for Feature Recognition of Curved Weld Seam on External Cylindrical Surface

In automated welding based on machine vision, the camera mounting angle has an important impact on recognizing weld seam features. In practical engineering applications, camera mounting angle deviation can lead to weld recognition confidence decay or algorithm failure, requiring high camera position mounting accuracy. Aiming at this technical bottleneck, this study proposes the HCSE-YOLO. To validate the performance of the HCSE-YOLO network architecture, two public datasets (Crack_Detection, Coco128) and a self-built dataset (Simulated_welds) are used. HCSE-YOLO improves the performance of the $\mathrm{mAP}_{50-95}$ over the base architecture, YOLOv10n, in the two public datasets by $5.7 \%$ and $5.68 \%$, respectively, and the occupied space size is reduced by $6.9 \%$ and $3.4 \%$, respectively. Meanwhile, trained on the self-built dataset, HCSE-YOLO improves by $4.5 \%$ and reduces the occupied space size by $5.26 \%$ over the base architecture $\mathrm{mAP}_{50-95}$. Experiments verify that HCSE-YOLO has high robustness to changes in camera deviation angle, and pictures taken at camera deviation angles of $0^{\circ} \sim 10^{\circ}$ are used as the training set. HCSE-YOLO can effectively identify weld features in images taken at camera deviation angles of $0^{\circ} \sim 40^{\circ}$. The improved model meets the requirements of real-time industrial production based on improving the accuracy of the identification of tiny gaps, enhances the generalization ability of the model, and broadens the options for the improvement of the YOLOv10 network architecture.