A welding defect detection network based on multi-dimensional collaborative enhancement of spatial-frequency features

Small diameter tubes are extensively applied in industries such as energy and petrochemicals. Poor welding quality potentially leads to serious safety hazards. Due to limited information available in X-ray images, and the variable scales and blurred edges of defects, accurate detection remains challenging. Most of existing methods of weld defect detection depend on spatial features. However, spatial domain features fail to effectively distinguish between the background and defects. To address these issues, we propose a welding defect detection network based on multi-dimensional collaborative enhancement of spatial-frequency features (MCSF-Net), which distinguishes defects from the background in the frequency-domain. The network comprises three key components: the multi-dimensional perception unit of defect features (MDPU), the spatial-frequency joint enhancement module (SFJE), and the aspect ratio-aware deep supervised (ARDS) head. MDPU captures differences between defect edges and background by combining different frequency sub-bands. SFJE reinforces the expression of detailed textures and semantic information of defect features. ARDS head balances the loss contribution of class-imbalanced defects by adjusting the sampling method for deep supervision and weighted loss. MCSF-Net improves mAP and recall by 2.9% and 6.7%, respectively, over state-of-the-art models.