Recognition of GTAW weld penetration based on the lightweight model and transfer learning

Wang, Zhenmin; Li, Liuyi; Chen, Haoyu; Lin, Sanbao; Wu, Jianwen; Ding, Tao; Tian, Jiyu; Xu, Mengjia

doi:10.1007/s40194-022-01396-0

Cited by 5 publications

(1 citation statement)

References 32 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Jiao et al [14] developed an algorithm based on Residual Neural Networks (ResNet) using transfer learning, which decreased the training time and improved the accuracy of weld penetration prediction. Wang et al [15] used the MobileNetV2-based transfer learning model to fit the custom dataset and recognize weld penetration states. Bahador et al [16] presents a novel application of transfer learning for tool wear detection in turning processes using onedimensional (1D) convolutional neural network (CNN).…”

mentioning

confidence: 99%

Arc bubble edge detection method based on deep transfer learning in underwater wet welding

Guo,

2024

Preprint

View full text Add to dashboard Cite

Arc bubble in underwater wet welding reflects the stability of the welding process. An arc bubble edge detection method based on deep transfer learning is proposed to overcome the shortcomings of conventional algorithms in processing underwater wet welding images. The method consists of two training stages: pre-training and fine-tuning. In the pre-training stage, a large source domain dataset is used to train VGG16 as a feature extractor. In the fine-tuning stage, we proposed the Attention-Scale-Semantics (ASS) model, which consists of a Convolutional Block Attention Module (CBAM), a Scale Fusion Module (SCM) and a Semantic Fusion Module (SEM). The ASS model is retrained with the small underwater wet welding target domain dataset to fine-tune the model parameters. The CBAM can adaptively weight the feature maps, focusing on more important feature to better capture edge information. The SCM training method makes extensive use of feature information to simplify the training steps. Additionally, the skip structure of SEM effectively resolves the problem of semantic loss in the high-level network during the training process and improves the accuracy of edge detection. We compare the ASS model to the conventional edge detection model on the BSDS500 dataset and underwater wet welding images, demonstrating that the ASS model is superior to the conventional edge detection model. By comparing with Richer Convolutional Features (RCF), Fully Convolution Network (FCN) and UNet, the excellent performance of the ASS model in arc bubble edge detection method is verified.

show abstract

mentioning

confidence: 99%