Confidence-aware agglomeration classification and segmentation of 2D microscopic food crystal images

2025-07-30

AI Summary

Accurate segmentation and classification of food crystal aggregates in 2D microscopic images remain challenging due to water-phase transparency and single-view limitations, leading to ambiguous boundaries and labeling difficulties. Method: We propose a joint classification-segmentation framework featuring a dual-model collaborative inference mechanism that integrates instance-level classification with pixel-level segmentation, augmented by a confidence-aware strategy to better discriminate transparent, adhesive water-phase regions. A morphology-preserving post-processing module is further designed to retain crystal structural characteristics and enhance prediction robustness. Supervised pseudo-label generation and end-to-end joint training enable effective learning under both high- and low-confidence annotation scenarios. Results: Our method significantly improves aggregate detection true positive rate and crystal size distribution estimation accuracy. It delivers an interpretable, high-precision automated solution for food microstructure analysis.

Technology Category

• Computer Vision: Segmentation
• Machine Learning: Multi-instance/Multi-view Learning
• Intelligent Robots: Multimodal Perception & Sensor Fusion

Application Category

• Search and Retrieval-Augmented AI: Agentic search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Security and Privacy: Data transparency and provenance

Abstract

Food crystal agglomeration is a phenomenon occurs during crystallization which traps water between crystals and affects food product quality. Manual annotation of agglomeration in 2D microscopic images is particularly difficult due to the transparency of water bonding and the limited perspective focusing on a single slide of the imaged sample. To address this challenge, we first propose a supervised baseline model to generate segmentation pseudo-labels for the coarsely labeled classification dataset. Next, an instance classification model that simultaneously performs pixel-wise segmentation is trained. Both models are used in the inference stage to combine their respective strengths in classification and segmentation. To preserve crystal properties, a post processing module is designed and included to both steps. Our method improves true positive agglomeration classification accuracy and size distribution predictions compared to other existing methods. Given the variability in confidence levels of manual annotations, our proposed method is evaluated under two confidence levels and successfully classifies potential agglomerated instances.

Problem

• Classify and segment 2D microscopic food crystal agglomerations
• Overcome manual annotation challenges due to water transparency
• Improve accuracy in agglomeration classification and size prediction

Innovation

• Supervised baseline model generates segmentation pseudo-labels
• Instance classification model with pixel-wise segmentation
• Post-processing module preserves crystal properties