Filtering Out Defects in Open-Source Meat Freshness Datasets: A New Leak-Free Benchmark and a Comparative Analysis of Four Different Mobile Deep Learning Models.

Image-based, nondestructive classification of hourly meat-freshness stages is important for food-quality monitoring and supply-chain management. However, detailed audits of open-source datasets and their preprocessed subsets can reveal methodological defects that artificially inflate model performance. In this study, a Mendeley Data dataset for hourly meat-freshness classification was audited at the file and original-image levels. Two major defects were identified: The 36- and 48-h beef classes contained byte-identical images, as verified by MD5 hashes, and offline augmentation performed before data partitioning distributed variants of the same original images across the training, validation, and test sets, causing cross-partition leakage. A leak-free nine-class benchmark containing 1849 verified original images was therefore constructed. The duplicated beef stages were represented as a single 36+ h beef class, whereas the distinct 36 and 48 h mutton stages were retained separately. MobileNetV3-Large, EfficientNet-B0, ShuffleNetV2, and MobileNetV2 were trained using online augmentation, focal loss, inverse-frequency class weights, and five prespecified random seeds. MobileNetV3-Large achieved the highest mean test accuracy (98.51% ± 0.58%) and macro F1 score (98.54% ± 0.55%) and ranked first in four of five runs. However, its advantage over EfficientNet-B0 was not significant in any seed-wise exact McNemar comparison after Holm correction. These findings demonstrate the risks of uncritical reuse of open-source datasets and provide a leakage-controlled, reproducible basis for evaluating compact models intended for mobile and edge deployment.