DRST-Net: A Dual-Branch Feature Fusion Network Combining ResNet50 and Swin Transformer for Welding Light Strip Recognition

To address the challenges of strong arc light noise, metal spatter, and smoke interference in weld seam recognition, we propose DRST-Net, a dual-branch cross-attention feature fusion network. The encoder of DRST-Net consists of two branches: ResNet50 for detailed feature extraction and Swin Transformer for global semantic modeling. The network also incorporates a cross-attention feature fusion (CaFF) module, which utilizes the self-attention mechanism to dynamically integrate local features from ResNet50 with global semantic information from Swin Transformer. This design allows DRST-Net to effectively merge both detailed and global features, enhancing the accuracy and robustness of weld seam recognition. The experimental results show that DRST-Net outperforms the current leading models, achieving a mIou of 86.17%, mPre of 92.32%, and F1_score of 91.68%. The model performs especially well in complex environments with noisy backgrounds and intricate details. Ablation experiments confirm the effectiveness of the dual-branch structure and the cross-attention fusion module. This approach provides a practical and reliable solution for automated welding tasks, offering significant improvements in both accuracy and robustness.