Benchmark-Driven Evaluation for Food Safety Vision Systems

Abstract

Food safety vision systems require benchmark-driven evaluation because models developed under narrow laboratory settings may fail when exposed to realistic variation in samples, sensors, illumination, and contamination levels. Hyperspectral imaging is particularly useful for meat inspection because it captures rich spectral information beyond visible appearance. However, weak contamination signals require evaluation protocols that measure sensitivity to subtle spectral mixtures rather than only classification accuracy. This topic focuses on benchmark-driven assessment of food safety vision systems, emphasizing dataset design, annotation reliability, spectral unmixing performance, and practical inspection robustness. Benchmarking also enables comparison among chemometric models, deep learning methods, matched filtering, and endmember-based approaches. The resulting literature direction supports reproducible food inspection research and practical deployment of non-destructive contamination detection systems.