Explainable AI-Driven Vision Transformers for Assessing Fruit Freshness via Transfer Learning

The assessment of fruit freshness is a critical factor in ensuring the quality and safety of food products, impacting both consumer satisfaction and market value. This paper presents a comparative evaluation of Vision Transformer (ViT) models for assessing fruit freshness using transfer learning techniques. Three pretrained ViT models—ViT model A (vit_base_patch16_clip_384), ViT model B (vit_betwixt_patch16_reg4_gap_384), and ViT model C (vit_mediumd_patch16_reg4_gap_384)—were employed on two publicly available datasets: Dataset A from the Mendeley Data Repository and Dataset B from the Kaggle Repository. Data augmentation techniques were applied to enhance the generalization capabilities of the models. Superior performance was exhibited by ViT model B on Dataset A, with an accuracy of 98.44%, precision of 98.46%, an F1-score of 98.44%, and a recall of 98.44%. Again, exceptional results were achieved by ViT model B on Dataset B, reaching an accuracy of 99.57%, precision of 99.56%, an F1-score of 99.56%, and a recall of 99.56%. The effectiveness of ViTs in accurately classifying fruit freshness is underscored, demonstrating their potential for practical applications in food quality assessment. Later, Local Interpretable Model-agnostic Explanation (LIME) was incorporated to improve the interpretability of the model.