BoT-YOLOv8: A high accuracy and stability initial weld position segmentation method for medium-thickness plate

Abstract

Abstract To address the technical bottleneck of autonomous vision guidance for the initial weld position of medium-thickness plate in robot welding. This paper proposes a high accuracy and stability initial weld position segmentation method for medium-thickness plate, this method is developed by integrating the Bottleneck Transformer (BoT) into YOLOv8, termed as BoT-YOLOv8. Firstly, aim to filter out redundant information in the image and enhance the model's capability to express features, the BoT is added behind the last bottleneck layer in the residual module of the YOLOv8 neck structure. Subsequently, in order to obtain the multi-scale information of the target, the atrous convolution is incorporated as the spatial pyramid pooling structure to establish connections between the backbone and the neck of this model. Furthermore, to facilitate the learning of weld position characteristics for the welding robot, the Hue-Saturation-Value (HSV) space region segmentation method is utilized to postprocess the weld seam features. Finally, ablation experiments are conducted on the self-created weld dataset. The results demonstrate that the proposed method achieves a trade-off between detection accuracy (93.1% \({mAP}^{0.5}\)) and detection speed (26.5 \(FPS\)) on a 12GB NVIDIA GeForce RTX 3060 GPU. In addition, compared with the existing methods, the presented method exhibits stronger anti-interference capability.