Comparison of MobileNetV2, EfficientNet-B0, and ResNet50 for Fruit Freshness Classification Based on Accuracy and F1-Score

Abstract

Freshness is a key determinant of fruit quality and safety, while manual assessment is subjective and slow. This study aims to compare three convolutional neural network architectures, namely MobileNetV2, EfficientNet-B0, and ResNet50, for the freshness classification of apples, bananas, and oranges in fresh and rotten conditions. The Food Freshness Dataset from Kaggle with 29,502 images was used and divided with a ratio of 70:20:10 into six classes. All three models were built using a transfer learning scheme with feature extraction on ImageNet pre-trained weights, an input size of 224×224, and an identical training configuration for 10 epochs. The main difference between the models lies in the specific preprocessing functions of each architecture to ensure a fair comparison. Evaluation was carried out on test data using accuracy and F1-score macro and weighted. The results show that ResNet50 achieved the highest performance with an accuracy of 0.9780 and a macro F1-score of 0.9787, followed by EfficientNet-B0 (0.9759; 0.9774) and MobileNetV2 (0.9726; 0.9732). Class-by-class analysis revealed that the Rotten Orange class was the most difficult for all models. EfficientNet-B0 and MobileNetV2 performed comparable to ResNet50 but with a much smaller number of parameters, making them more efficient for resource-constrained applications. This study emphasizes the importance of reporting F1-scores alongside accuracy on class-imbalanced data.