A Dual-Dataset Study on Deep Learning-Based Tropical Fruit Classification

This work uses two datasets to investigate deep learning models for fruit classification: a proprietary dataset containing photos of eight different varieties of fruit from India, and the publicly accessible FIDST-30. The YOLOv3 and YOLOv7 models were employed for object detection, while for picture classification, ResNet50 and VGG16 were utilized. On the FIDST-30 dataset, VGG16 outperformed ResNet50 with a mean F1 score of 0.878, which was higher than its previous score of 0.843. Methods for augmenting data enhanced the resilience of the model. ResNet50 and VGG16 demonstrated domain adaptation tests with testing accuracies of $85 \%$ and $86 \%$, respectively. The outcomes demonstrate the superior performance of VGG16 and the advantages of domain adaption and data augmentation in raising fruit classification accuracy.