GPESNet: Global Prompt-Embedded Segmentation Network for Weld Defect Detection

Accurate defect detection in weld seams is crucial for ensuring the structural integrity of industrial components. Deep learning methods like Transformer-based networks have made promising achievements in natural image processing, providing potential solution for weld defect segmentation. However, due to the diverse and complex nature of weld defects, existing methods still face the challenge in effectively incorporating global contextual information in defect segmentation. To address this limitation, this paper proposes Global Prompt-Embedded Segmentation Network (GPESNet), which enhances global awareness by introducing a global prompt into the Transformer decoder. By enabling each query feature to embed global context information, our model enhances feature representation and improves segmentation accuracy. To evaluate our method, we constructed a weld defect segmentation dataset composed of real-world X-ray images with diverse defect morphologies. Experimental results show that our method outperforms other methods in weld defect segmentation, achieving a mIoU of 82.22% and a mPA of 91.01 %. These results demonstrate the effectiveness of GPESNet in industrial defect analysis.