Lightweight defect detection algorithm for wire and arc additive manufacturing based on modified YOLOv8 model

Surface defects that affect the quality of parts are a particularly significant issue in wire arc additive manufacturing processes (WAAM). Therefore, how to effectively control their surface quality has become a focus of researchers' attention. However, due to limited computing power and storage space of terminal devices, it is difficult to deploy defect detection models. Therefore, We present a lightweight WAAM weld surface defect detection algorithm based on YOLOv8n, called high-alternative novel YOLO (HAN-YOLO). Specifically, a novel lightweight adaptive Inverted bottleneck (NLAIB) is designed to optimize lightweight network architectures while significantly improving inference speed and computational efficiency. Subsequently, a lightweight alternative alterable kernel convolution (AAKConv) is employed to improve detection accuracy while reducing model parameters and complexity. Furthermore, the High-Level Screening Feature Fusion Pyramid (HS-FPN) was integrated to achieve multi-scale object detection, enhancing the model's feature selection and fusion capabilities. Finally, experiments on the 3440-WAAM weld surface defect dataset, NEU-DET dataset and Weld dataset are made to test the validity of HAN-YOLO. The experimental results show that, compared with YOLOv8n, the model parameters and GFLOPs of HAN-YOLO are reduced by 44.1% and 39%, respectively. Moreover, HAN-YOLO achieves an increase of 1%, 6.3%, and 38.6% in mAP@0.5, mAP@0.5:0.95, and real-time detection speed (FPS), respectively. These results demonstrate that HAN-YOLO is effective, and provides a lightweight detection scheme for the weld defects in WAAM.