Defect Detection for Steel Pipe Welding Defect Based on TransUNet

Automatic detection technology for steel pipe weld defects is an important and challenging task in the field of computer vision. Due to the irregularity of the shape, size, distribution area, color and appearance of welding defects, as well as the complex background, occlusion and noise, etc., weld defect detection remains an important and challenging issue. U-Net has become the de facto standard for a variety of complex image segmentation tasks and has achieved great success. However, due to the inherent locality of convolution operations, UNet generally exhibits limitations in explicitly modeling remote dependencies. Transformers are designed for sequence-tosequence prediction and have emerged as an alternative architecture with an inherent global self-attention mechanism, but the lack of low-level details can lead to limited localization capabilities. This article proposes an TransUNet method for detecting defects in steel pipe welds. This method innovatively builds upon the classic U-Net architecture by introducing the Transformer module to enhance the global modeling capability of the model. It utilizes multi-scale dilated convolutions to extract weld seam features at different scales and dynamically focuses on key regions in the image through visual attention mechanisms. In the decoding stage, cascaded upsampling is used to restore image resolution, and skip connections are used to avoid information loss. Experiments were conducted on the steel pipe welding defect image database, and the results showed that the proposed method has a more accurate segmentation effect in steel pipe weld defect detection, with a detection accuracy rate of up to 92.23 %. The detection accuracy is high, and the model has strong recognition ability for complex scenes and subtle structures.