Abstract
Bubble detection in gypsum board inspection remains challenging, as many existing methods rely on large-scale annotated datasets that are costly to obtain and may limit practical use in low-annotation industrial settings. To address this limitation, we propose DiffuSeg, a reference-guided and training-free framework for bubble detection and segmentation in gypsum board images. Given a single box prompt, the method first extracts a reference bubble mask and constructs a global similarity map to identify candidate regions. To handle bubbles with weak direct similarity to the initial reference, we introduce a Dual-Path Similarity Diffusion (DPSD) module, which iteratively updates the reference representation by aggregating features from high-confidence target regions. Meanwhile, a parallel noise path models distracting responses, enabling cross-path interaction to enhance true bubble instances while suppressing false positives and mitigating semantic drift. To further improve boundary localization, we propose Contrastive Peak-Guided Segmentation (CPGS), which generates paired positive and negative prompts from target and noise peaks, respectively, and feeds them into SAM to produce refined instance masks. Experiments on a self-built gypsum board bubble dataset show that DiffuSeg achieves an AP of 58.8\%, outperforming the best-performing baseline by 6.2 percentage points, while reducing the counting MAE by 9.7. These results validate the effectiveness of DiffuSeg on the self-built gypsum board bubble dataset and suggest its potential usefulness for bubble measurement under low annotation and training requirements.
Funding Information
  • Leading Goose" R&D Program of Zhejiang Province (2024C01028)
  • State Key Laboratory of Industrial Control Technology, China (ICT2025C02)

This publication has 17 references indexed in Scilit: