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

Area of Science:

• Welding Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Underwater wet welding stability relies on arc bubble characteristics.
• Existing arc bubble edge detection methods yield blurry and discontinuous results.
• Limited research addresses arc bubble edge detection in underwater welding.

Purpose of the Study:

• To propose a novel deep transfer learning method for accurate arc bubble edge detection in underwater wet welding images.
• To develop an Attention-Scale-Semantics (ASS) model incorporating CBAM, SCM, and SEM modules.
• To evaluate the ASS model's performance against traditional and state-of-the-art edge detection techniques.

Main Methods:

• Employs a two-stage deep transfer learning approach: VGG16 pre-training and ASS model fine-tuning.
• The ASS model integrates Convolutional Block Attention Module (CBAM), Scale Fusion Module (SCM), and Semantic Fusion Module (SEM).
• Evaluated on BSDS500 and a custom underwater wet welding dataset, comparing against RCF, FCN, UNet, LDC, and TEED.

Main Results:

• The ASS model demonstrates superior performance in Mean Absolute Error (MAE) and accuracy compared to benchmark models.
• Adaptive feature weighting by CBAM enhances crucial edge information capture.
• SCM and SEM modules effectively utilize multi-scale features and mitigate semantic loss, improving detection accuracy.

Conclusions:

• The proposed ASS model offers effective and stable arc bubble edge detection for underwater wet welding images.
• Deep transfer learning provides a robust framework for addressing challenges in specialized image analysis.
• The ASS model represents a significant advancement in monitoring and ensuring the quality of underwater welding processes.