ARCHIVES

This paper presents MeatScan v4, a high-resolution deep learning system for automated binary classification of meat freshness (Fresh vs. Spoiled) from smartphone-captured images. Built on a ConvNeXt-Base backbone pre-trained on ImageNet-1K and fine-tuned via a two-phase transfer learning protocol, the model is trained on 10,517 high-resolution images (3024×4032 px) captured from two Ghanaian markets. The system achieves 100% test accuracy, AUC-ROC of 1.0000, and zero false negatives under multi-scale Test-Time Augmentation (TTA) across three spatial scales, a conservative decision threshold of τ = 0.40 for food safety sensitivity, and gradient accumulation enabling training on a consumer-grade GPU. Ablation experiments confirm that both TTA and two-phase training are necessary for peak performance. Comparison against ResNet-50, EfficientNet-B0, and MobileNetV3 baselines on the same dataset demonstrates the superiority of the proposed approach. These findings validate deep learning-based meat inspection as a practical, scalable alternative to manual visual assessment.