Single-dataset orange image benchmark reports 99.12% accuracy with ViT, PCA

A leakage-controlled benchmark study using a single orange-fruit dataset reports 99.12% mean classification accuracy with a ViT + PCA + SVM pipeline. Researchers tested 32 combinations built from eight fixed feature extractors—seven CNNs plus a ViT-S/16 Vision Transformer—and four classical classifiers (SVM, k-nearest neighbors, random forest and XGBoost). The top pipeline reached 99.12% ± 0.71% over 25 cross-validation partitions, beating the strongest CNN setup, DenseNet201 + PCA + SVM, at 98.48% ± 0.81%. Evaluation used 5×5 repeated stratified cross-validation with an 80:20 class-stratified train-test split, and PCA was fit only on training folds. Results showed SVM dominance after PCA reduction and F1 scores below 98% for Black Spot and Canker versus near-perfect performance on Fresh and Greening.