Research on robot vision groove recognition algorithm based on improved YOLOv8

• Published in: International journal on interactive design and manufacturing Vol. 19; no. 11; pp. 7777 - 7789
• Main Authors: Gao, Shan, Dong, Bing, Li, Youhao, Li, Yajie
• Format: Journal Article
• Language: English
• Published: Paris Springer Paris 01.11.2025; Springer Nature B.V
• Subjects: Accuracy; Algorithms; Annotations; Automatic welding; CAE) and Design; Cameras; Classification; Coal mining; Computer-Aided Engineering (CAD; Datasets; Deep learning; Efficiency; Electronics and Microelectronics; Emitters; Engineering; Engineering Design; Formability; Grooves; Industrial Design; Instrumentation; Laser beam welding; Machine learning; Machine vision; Mechanical Engineering; Methods; Original Article; Recognition; Robots; Vision systems; Deep learning; Machine vision; Groove recognition; YOLOv8 model
• ISSN: 1955-2513; 1955-2505
• DOI: 10.1007/s12008-025-02342-2

To address the issue of groove recognition in robotic automatic welding, an improved groove detection algorithm based on YOLOv8 is proposed in this paper. The algorithm enables the classification and detection of seven common types of grooves without needing a laser emitter. By introducing the FasterNet Block into the Backbone and using ParitalConv to reduce redundant computations and memory access, the algorithm improves computational speed without sacrificing accuracy. Additionally, the incorporation of the Deformable Attention Transformer (DAT) in the Neck part allows the model to focus more on groove contours, effectively integrating low-level detail features with high-level semantic features. The improved YOLOv8n-FasterNet-DAT model achieves an mAP@0.5 of 96.17%, representing a 3.21% improvement over the original YOLOv8n model. Furthermore, the Precision and Recall reach 96.82% and 94.41%, respectively. The enhanced model is characterized by high accuracy, short processing time, and strong stability in groove type recognition. Graphical Abstract