From Few Images to High Accuracy: Augmentation and Embedding Methods for Date Fruit Ripeness

Manual date harvesting and sorting remain labor-intensive and error-prone, particularly when distinguishing intermediate ripeness stages such as Rotab. We present an image-based classification pipeline for the Berhi cultivar that assigns fruit to three ripeness stages—Khalal, Rotab, and Tamar—using compact deep structures and training strategies suited to small datasets. Rather than relying on generative or adversarial methods, our approach emphasizes (i) careful augmentation (classical transforms, automated policies, and sample-mixing), (ii) transfer and self-supervised pre training, and (iii) embedding- and metric-learning alternatives, with ensembles and test-time augmentation used as optional accuracy/robustness boosters. On a ۱۵۰-image dataset (۵۰ images per class) evaluated with ۵-fold cross-validation, a ResNet۱۸ baseline reaches about ۹۵% average accuracy. Automated augmentation combined with MixUp/CutMix improves accuracy to ۹۷%, and self-supervised pre training plus advanced augmentation and ensembling attain peak performance near ۹۸%. Improvements are most pronounced for the visually ambiguous Rotab class. We also report practical robustness measures (common corruptions, geometric stability, and calibration), which show that augmentation and pre training substantially increase stability under realistic input variability. These results indicate that, for small and visually subtle datasets, augmentation and pre training—rather than synthetic data generation—offer a pragmatic path to high accuracy and robust behavior.

کلیدواژه ها:

Date fruit ، Ripeness ، Deep Learning ، Self-supervised learning ، Metric learning ، Robustness