BoT-YOLOv8: a highly accurate and stable initial weld position segmentation method for medium-thickness plate
摘要
Due to the complexity of the initial welding position types in medium-thickness plate workpieces, the welding robot struggles to achieve automatic vision guidance before welding these workpieces. To address this problem, an initial welding position segmentation method with high accuracy and stability for medium-thickness plate is proposed. This method is developed through the integration of the Bottleneck Transformer (BoT) with YOLOv8, enabling a novel framework termed BoT-YOLOv8. To optimize the filtration of redundant information in the image and enhance the model’s feature representation ability, the Bottleneck Transformer is incorporated following the final bottleneck layer within the residual module of the YOLOv8 neck structure. Subsequently, to capture multi-scale information of the target features, the atrous convolution is integrated into the spatial pyramid pooling structure. Furthermore, to enhance the learning of welding position characteristics for the robotic welding system, the hue-saturation-value (HSV) color space region segmentation method is employed to post-process the weld seam features. Finally, ablation experiments are conducted on the self-created weld dataset. The results demonstrate that the proposed method strikes a trade-off between detection accuracy (93.1%