PBDWNet: a multi-scale porcelain surface bubble segmentation network based on frequency-domain fusion and deformable feature awareness

Microscopic bubbles within porcelain glazes encapsulate critical parameters of ancient ceramic production techniques. These bubbles are small, low-contrast, and poorly defined against complex backgrounds, making detail preservation and noise suppression challenging for existing methods. To address these limitations, this study employs a three-branch PBDWNet to extract the intrinsic structural information of bubbles. In the structure-preserving pathway, deformable feature awareness integrates deformable convolution with channel attention to adaptively capture bubble morphology and enhance multi-scale feature interaction. The detail enhancement branch applies wavelet feature fusion to decompose images into high- and low-frequency components, strengthening edge textures while suppressing background noise. The backbone network achieves unified representation of global and local information through cross-scale feature fusion. Experiments on the PRMI dataset show that PBDWNet outperforms mainstream segmentation models, surpassing the second best by 0.79% in bubble IoU and achieving the highest mean F1-score of 91.70%, demonstrating superior accuracy in complex scenarios.