Related Experiment Video
Updated: May 8, 2025
04:48
Swin-PSAxialNet: An Efficient Multi-Organ Segmentation Technique
Published on: July 5, 2024
318
CvT-UNet: A weld pool segmentation method integrating a CNN and a transformer
Longcheng Yang1,2, Huajun Wang2, Wenjie Meng2
1Sichuan Key Laboratory of Indoor Space Layout Optimization and Security Guarantee, Chengdu Normal University, Chengdu, 611130, China.
Heliyon
|February 17, 2025
Summary
This study introduces CvT-UNet, a novel semantic segmentation model for weld seam defect identification. The model effectively combines Transformer and Convolutional Neural Network strengths for improved fusion pool segmentation, achieving high accuracy with fewer parameters.
Area of Science:
- Computer Vision
- Machine Learning
- Materials Science
Background:
- Weld seam defect identification is challenging due to weak edges caused by light and impurity interference.
- Segmentation of the weld seam fusion pool is crucial for accurate defect detection.
- Transformers offer superior sequence modeling but lack local information extraction, unlike Convolutional Neural Networks (CNNs).
Purpose of the Study:
- To develop a novel semantic segmentation model, CvT-UNet, that integrates the global context capabilities of Transformers with the spatial information advantages of CNNs.
- To address the limitations of existing models in weld seam image segmentation, particularly in challenging environmental conditions.
- To improve the precision and efficiency of weld seam fusion pool segmentation for defect identification.
Main Methods:
- A novel CvT-UNet semantic segmentation model is proposed, utilizing an encoder-decoder architecture with designed CvT blocks.
- The model combines Transformer's self-attention mechanism for global context with CNN's local feature extraction.
- Depthwise separable convolutions and a redesigned skip-connection module are employed to optimize parameter usage and segmentation accuracy.
Main Results:
- The CvT-UNet model achieved average Intersection over Union (IoU) scores of 93.75%, 88.31%, and 90.86% on diverse weld seam datasets.
- On an automotive weld seam dataset, CvT-UNet outperformed UNet3+ by 1.63% in mean IoU (MIOU) and 1.39% in mean pixel accuracy (MPA).
- CvT-UNet also showed improvements over TransUNet and LIOT, demonstrating superior segmentation performance with fewer parameters.
Conclusions:
- The proposed CvT-UNet model effectively combines Transformer and CNN features for precise weld seam fusion pool segmentation.
- The model demonstrates significant improvements in segmentation accuracy and efficiency compared to existing state-of-the-art methods.
- CvT-UNet offers a promising solution for robust weld seam defect identification in challenging industrial environments.
