TEGDNet: Texture Enhancement Guided Detection Network for spiral welded pipeline defect detection

Spiral welded pipelines play a crucial role in the transportation industry. However, defects in the pipeline welds pose significant safety hazards, and their detection is challenging. Current detection methods are highly subjective, labor-intensive, and prone to missed or overlooked defects without sufficient texture information. Therefore, we propose a solution to address these challenges. Firstly, we construct a dataset of spiral welded pipeline weld defects (NEU-WELD-2000). Furthermore, we propose a texture enhancement guided detection network (TEGDNet) to detect defects with insufficient texture information, such as weak textures, strong interference, scale variation, and tiny targets. TEGDNet includes an encoder structure for feature extraction and enhancement and a decoder structure based on super-resolution reconstruction. Experiments have demonstrated that our method achieves excellent detection results with fewer than 4M parameters, while mAP₅₀ improves by 3.2%, and mAP₇₅, which is typically harder to improve, increases by 1.2% compared to the baseline. Additionally, we validate the effectiveness of each module through ablation experiments. Finally, the proposed detection method shows great potential in actual pipeline weld detection, significantly saving time and costs for operators through visualization software. The source code and the dataset are publicly available at https://github.com/VDT-2048/TEGDNet.