Fruit freshness prediction using transfer learning and image classification

Abstract

Accurate detection of fruit freshness is vital for quality control, spoilage reduction, and supply chain efficiency. This study introduces a classification system using transfer learning with a pre-trained ResNet-18 on 55,824 images of fresh and rotten fruits across 10 classes. Extensive augmentation improved generalization, and training with PyTorch and GPU acceleration employed early stopping and checkpointing to avoid overfitting. Evaluation metrics (precision, recall, F1-score, accuracy) show consistently high performance, with smooth loss convergence and a confusion matrix confirming reliable distinction between visually similar fruits. Overall, transfer learning with ResNet-18 proves accurate and efficient for automated fruit freshness classification