Leveraging pre-trained computer vision models for accurate classification of meat freshness

Abstract

Increasing concerns about food quality and safety have led to research into ways to assess meat freshness. Advances in deep learning, particularly image classification, enable up new possibilities for fast and non-destructive methods of evaluating meat properties. This study explored a novel approach for classifying meat freshness based on image data by leveraging features extracted from pre-trained deep convolutional neural networks (DCNNs), followed by random encoding of aggregated deep activation maps (RADAM). The encoded features were subsequently used to train traditional machine learning (ML) classifiers. This approach yielded state-of-the-art results, with classification metrics ranging from 93 to 100 % when classifying beef and chicken meat freshness across three datasets and under multiple hyperparameter settings. Not only does this surpass the performance reported in literature, but it also offers a simpler and more efficient methodology. Findings suggest that the proposed approach could be a practical and effective solution for industry deployment.