A welding defect detection method based on multiscale feature enhancement and aggregation
Lukui Shi, Shiyuan Zhao, Weifei Niu
Abstract
Welding defects can pose significant safety risks in the industrial production. Accurate detection of these defects is crucial for ensuring the industrial production safety. Traditional computer vision methods rely on operator experience and knowledge and are time-consuming. In addition, manual feature extraction and classifier definition become more difficult for complex images or unknown defects. In contrast, deep learning can automatically extract features from raw data and build classifiers with higher accuracy. Additionally, deep learning models have stronger generalization ability and can identify a wider range of defects, without being limited by feature engineering. Consequently, we proposed a welding defect detection model based on multiscale feature enhancement and aggregation based on the advanced deep learning model YOLOv7. The model is composed of a wavelet multiscale attention (WMA), a parallel sampling module (PSS), and a path expansion aggregation network (PEANet). We compared the model with a variety of excellent object detection algorithms. Experiments show that the proposed model can effectively identify welding defects. Compared with the original model, the mAP has increased by approximately 1.2%, the average recall has increased by approximately 5.3%, the parameter volume has increased slightly, and the Flops has decreased by approximately 62.6%..