Smartphone-Assisted Aerogel Colorimetric Sensor Array Coupled with Lightweight Deep Learning for Multi-Meat Freshness Monitoring

Rapid assessment of chilled-meat freshness is constrained by conventional laboratory-based measurements. Although food-detection studies emphasize predictive performance, computational cost is reported less consistently. Here, we developed a 2 × 2 colorimetric sensor array based on a porous konjac glucomannan (KGM)/gelatin/starch/SiO2 aerogel substrate loaded with four pH-sensitive dyes. Sensor images were acquired with a smartphone under standardized conditions, and four dye-responsive regions of interest (ROIs) were manually extracted and assembled into 2 × 2 dye-grid images. Five lightweight models—Spot-CNN, ShuffleNetV2-0.5x, EfficientNet-Lite0, MobileNetV3-Small and Tiny-MLP—were evaluated for classifying beef and pork freshness as Fresh, Sub-fresh or Spoiled. During chilled storage, the sensor color responses showed broadly consistent temporal trends with total volatile basic nitrogen (TVB-N). Under the fixed group-level split, Spot-CNN achieved image-unit-level accuracies of 0.933 and 0.938 and macro-AUC values of 0.995 and 0.982 on the beef and pork test sets, respectively. Tiny-MLP contained 0.013 M parameters and required less than 0.001 G floating-point operations, demonstrating lower computational cost. These results establish a controlled-laboratory prototype integrating an aerogel colorimetric sensor array, standardized smartphone image processing and lightweight-model comparison for multi-meat freshness screening. Further validation across meat batches, smartphone devices, lighting conditions and deployment environments is required.