Improved YOLOv5-based pore defect detection algorithm for wire arc additive manufacturing

The identification and detection of microstructural defects in wire arc additive manufacturing (WAAM) specimens play a significant role in characterizing and analyzing the mechanical properties of these specimens. Porosity is one of the primary forms of microstructural defects in WAAM specimens. Therefore, in this paper, a model for detecting pore defects in the cross-section of WAAM specimens based on the YOLOv5s model is proposed. The proposed model incorporates several improvements to enhance its speed and accuracy. Firstly, a lightweight backbone network is constructed to improve the detection speed by introducing the GhostConv and PConv modules. Secondly, an efficient spatial pooling pyramid structure (ESPP) is designed to enhance model processing speed and detection accuracy. Thirdly, the neck network incorporates a newly designed double path aggregation network (DPAN) to enhance the preservation of intricate details within the network. Finally, the C3 module in the YOLOv5s feature fusion network is improved, and the GC3ECA module is proposed by combining the efficient channel attention mechanism ECA and GhostConv module to enhance channel information and reduce redundant information. Tests on the self-built WAAM defect dataset and NEU-DET dataset show that the mean average precision(mAP)of the improved model is 90.4% and 76.5%, which is 1.9% and 3.8% higher than that of the YOLOv5s model, respectively. Meanwhile, the detection speed of the improved model reaches 74.21 frames per second (FPS), which is 38.7% higher than YOLOv5s.The experimental results show that the improved model has better overall performance.