Food Freshness Evaluation Using a CLIP-Based Architecture

In this work, we present an efficient deep learning framework for automated fresh and stale food classification using transfer learning with a pretrained CLIP-based feature extractor. The proposed system employs frozen vision transformer (ViT) embeddings from CLIP as generalized visual descriptors and integrates them with a lightweight multi-layer perceptron (MLP) classifier for binary classification. To enhance generalization, extensive data augmentation and stratified dataset partitioning were applied to the publicly available Fresh and Stale Classification dataset. Experimental results reveal a consistent improvement across ten training epochs, achieving a final test accuracy of 97.99%, F1-score of 0.9808, and ROC–AUC of 0.9985. The proposed model demonstrates excellent discriminative performance, robust convergence, and strong generalization capabilities while maintaining computational efficiency. These results confirm the suitability of CLIP-based visual representations for high-accuracy food quality assessment and real-time freshness detection applications.