Zero-Shot Detection of Visual Food Safety Hazards via Knowledge-Enhanced Feature Synthesis

Automated food safety inspection systems rely heavily on the visual detection of contamination, spoilage, and foreign objects in food products. Current approaches typically require extensive labeled training data for each specific hazard type, limiting generalizability to novel or rare safety issues. We propose a zero-shot detection framework for visual food safety hazards that enables the identification of previously unseen contamination types without requiring explicit training examples. Our approach adapts and extends the Knowledge-Enhanced Feature Synthesizer (KEFS) methodology to the food safety domain by constructing a specialized knowledge graph that encodes visual safety attributes and their correlations with food categories. We introduce a Food Safety Knowledge Graph (FSKG) that models the relationships between 26 food categories and 48 visual safety attributes (e.g., discoloration, mold patterns, foreign material characteristics) extracted from food safety databases and expert knowledge. Using this graph as the prior knowledge, our system synthesizes discriminative visual features for unseen hazard classes through a multi-source graph fusion module and region feature diffusion model. Experiments on our newly constructed Food Safety Visual Hazards (FSVH) dataset demonstrate that our approach achieves 63.7% mAP in zero-shot hazard detection, outperforming state-of-the-art general zero-shot detection methods by 6.9%. Furthermore, our framework demonstrates robust generalization to fine-grained novel hazard categories while maintaining high detection performance (59.8% harmonic mean) in generalized zero-shot scenarios where both seen and unseen hazards may occur simultaneously. This work represents a significant advancement toward automated, generalizable food safety inspection systems capable of adapting to emerging visual hazards without a costly retraining process.