A two-stage decoupled visual framework integrated with multi-indicator colorimetric sensor arrays for edge-computed meat freshness monitoring.

Developing highly precise and economically viable systems for the continuous monitoring of meat freshness is critical for mitigating food safety risks. However, conventional colorimetric sensor arrays are frequently limited by their reliance on a single indicator and susceptibility to complex background interference during edge deployment. To overcome these limitations, herein, we propose an AI-integrated, two-stage decoupled visual framework coupled with a multi-indicator colorimetric sensor array for robust meat freshness evaluation. A cross-reactive matrix integrating genipin, anthocyanin, and Cu-PAN probes into a polyvinyl alcohol/glycerol substrate enables the concurrent tracking of biogenic amines, pH changes induced by volatile basic amines, and hydrogen sulfide. To eliminate subjective decoding ambiguity and visual noise from heterogeneous packaging, a highly efficient edge computing architecture was developed. In the primary localization stage, an optimized You Only Look Once (YOLO) v11n architecture featuring a custom C3k2SGP module autonomously isolates the sensor array, compressing the parameter count by 36.5% to merely 1.64 M. Subsequently, a lightweight MobileNetV4 network, augmented with a novel Spatial Variance Coordinate Attention (SVCA) module, decodes the isolated colorimetric signals. By explicitly modeling localized spatial heterogeneity, this decoupled framework achieves an outstanding classification accuracy of 99.53% across four freshness grades. Seamlessly integrated into an edge graphical user interface with a 2.05 s inference latency, this portable platform provides a highly reliable and scalable solution for intelligent food packaging.