摘要
Abstract
The precise semantic segmentation of the welding pool is a key technology for achieving automated quality moni-toring of the welding process.Traditional recognition methods are constrained by arc light interference,dynamic target posi-tioning deviation,and insufficient feature fusion,making it difficult to meet the pixel-level segmentation requirements under complex conditions.A semantic segmentation model of the welding pool based on an improved YOLOv8s is proposed.Firstly,a CA attention mechanism is introduced into the backbone network,and the spatial information embedding and atten-tion generation process are separated to enhance the model's accurate positioning ability of the welding pool,side walls,and welding wire.Secondly,to address the loss of positioning information in the multi-scale feature fusion of the original PANet feature pyramid,a Bidirectional Feature Pyramid Network(BiFPN)is used as a replacement,and a dedicated segmentation layer is added.Through bidirectional paths,the interaction and fusion of features at different levels are strengthened,improv-ing the segmentation capability of the details at the edge of the welding pool.Finally,the EIoU loss function is introduced as a replacement for the original CIoU.By optimizing the predicted and actual boxes'length,width matching degree,and over-lap area evaluation separately,the positioning deviation of dynamic welding pool targets is reduced,enhancing the model's adaptability to changes in the size of the welding pool.Experimental results show that the improved model can achieve high-precision semantic segmentation of the welding pool and its surrounding areas,with a recognition accuracy of 99.37%and an intersection over union(IoU)of 93.76%,significantly improving the recognition effect.This research provides a reliable visual perception solution for intelligent monitoring of the welding process,which has significant engineering significance for promoting the upgrading of welding automation technology.关键词
熔池识别/YOLOv8s/CA注意力机制/语义分割Key words
molten pool identification/YOLOv8s/CA attention mechanism/semantic segmentation分类
金属材料
