Welding-sidewall spatter defect detection based on ECBAM-YOLO

Industrial weld splatter defects on sidewalls are typically caused by high temperatures, excessive solder, or improper operations, which severely affect equipment performance and increase the risk of structural failure. To address this issue, this paper proposes a defect detection method that integrates multispectral light source optimization with deep learning-based feature enhancement. The method employs a multispectral imaging system to capture sidewall regions of varying morphologies, which are then radially projected and segmented into small-sample images, enabling adaptive segmentation and classification. Subsequently, a geometric morphology-based synthesis approach is used to expand the defect dataset. Simultaneously, dynamic noise injection is applied to regions of interest within the samples, enhancing the diversity and realism of the data. This forms a physics-constrained, two-stage data augmentation strategy, which significantly improves the accuracy of defect feature detection. Finally, by enhancing channel responses and spatial focus, a detection network named ECBAM-YOLO is designed, which integrates a channel–spatial mixed attention mechanism to strengthen the semantic discrimination of defect texture features. Experimental results demonstrate that the proposed method exhibits excellent adaptability and precision in sidewall weld defect detection. It can effectively detect defects of various sizes and heights, showing strong potential for industrial applications.