Prototypical Modal Rebalance-Based Heterogeneous Sensor Fusion for Non-Destructive Pork Freshness Detection

Pork, as the main source of animal protein, accounts for approximately 34% of global meat consumption and requires reliable freshness assessment to ensure safety and quality. Traditional methods, such as electrical impedance spectroscopy and microbial analysis, can effectively evaluate meat freshness but often suffer from destructiveness, complex equipment, or low efficiency. While non-destructive single-modality sensors (e.g., RGB cameras or gas sensors) have emerged as alternatives, they are inherently constrained by environmental susceptibility and incomplete feature representation. Integrating visual and olfactory signals provides a promising solution; however, fusing heterogeneous sensor data frequently triggers a ”modal imbalance” problem, where the dominant modality suppresses the weaker one during deep learning optimization. To address this engineering challenge, this study develops an engineering-oriented heterogeneous sensor data fusion prototype, integrating the Prototypical Modal Rebalance (PMR) strategy for non-destructive pork freshness detection. First, we adapt an extended Adaptive Multi-Scale Attention U-Net (AMSAU-Net) as a visual pre-processing module. Acting as a hard attention mechanism, it precisely segments lean meat regions to filter background noise and extract high-purity visual features. Furthermore, the PMR strategy is introduced to dynamically balance the learning dynamics between the visual and olfactory modalities, maximizing their complementary benefits. Experimental results demonstrate that the proposed multimodal system achieves an outstanding accuracy of 99.58% across three freshness levels based on Leave-One-Subject-Out Cross-Validation (LOSOCV) using storage-time proxy labels, while TVB-N measurements are mainly utilized for external validation and biochemical calibration. Comprehensive ablation studies across multiple classifiers confirm that the PMR strategy significantly improves the recall rate by effectively compensating for the weaker olfactory signals, providing a highly accurate and reproducible prototypical solution for the freshness classification of cut pork samples under controlled conditions.