UQ-SegNet: An instance segmentation method with uncertainty quantification for crystallization processes
Abstract Reliable crystal detection is critical for crystallization quality control. Small, densely distributed crystals dominate crystallization images. Their blurred boundaries and low contrast make segmentation prone to missing them or merging neighboring crystals. These errors distort the crystal size distribution (CSD) obtained from the segmentation results, so segmentation reliability must also be assessed. We propose UQ-SegNet to segment individual crystals and quantify pixel-level uncertainty. A High-Frequency Fusion Multiscale Convolutional Attention module combines discrete cosine transform filtering with multiscale attention to retain small-crystal edges. A Position-Aware Head adds positional encoding to ROI features to improve crystal localization, and a Bayesian Decoder estimates epistemic and aleatoric uncertainty with Monte Carlo Dropout. UQ-SegNet achieves the highest AP75 (72.13%) and APs (61.30%) among compared models, and its CSD agrees most closely with the annotations. Both uncertainty types correlate positively with pixel errors. The method provides the CSD with pixel-level uncertainty for assessing the underlying segmentation.
Authors
- Liangyong Wang (ORCID: https://orcid.org/0000-0001-7811-5228)
- Peng Fu (ORCID: https://orcid.org/0009-0001-8101-4468)
- Bohua Hu
Institutions
- Northeastern University (US)
Publication Details
- Journal
- Measurement Science and Technology
- Published
- 2026-11-17
- DOI
- https://doi.org/10.1088/1361-6501/aec318
- Primary Topic
- Advanced Neural Network Applications
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- National Natural Science Foundation of China