Deep Learning-Based Beef Freshness Classification: A Comparative Analysis of Six CNN Architectures with Transfer Learning

Abstract

The freshness of meat products is important for food safety and consumer health. Manual inspection methods are subjective and hard to scale, which leads to the need for automated vision-based solutions. This study compares six pretrained convolutional neural network architectures which are MobileNetV2, MobileNetV3, EfficientNetB0, ResNet50, DenseNet121, and InceptionV3 for classifying beef freshness into three categories which are Fresh, Half-Fresh, and Spoiled. All models were trained using a two-phase transfer learning strategy that includes feature extraction and selective fine-tuning, applied to a dataset of 2266 beef images that were sourced from Roboflow Universe. Each model was initialized with ImageNet weights and evaluated under the same experimental conditions. Experiments were done on an NVIDIA Tesla T4 GPU and used TensorFlow 2.19. Model performance was assessed with metrics such as test accuracy, macro and weighted F1-score, precision, recall, and exported model size. Results show that MobileNetV2 reached the highest test accuracy of 99.12% and a macro F1-score of 99.17%, and it also had the smallest model size of 21.26 MB compared to all other evaluated architectures. These results suggest that MobileNetV2 is the most effective and computationally efficient architecture for classifying beef freshness, which makes it suitable for practical use in automated food quality inspection systems.