Benchmarking Deep Learning Models for Visual Classification and Segmentation of Horticultural Commodities

Recent advances in computer vision have enabled new approaches for automated quality assessment of tropical fruits, where accurate classification and segmentation are essential for postharvest inspection. A major challenge lies in identifying deep learning architectures that achieve high accuracy while remaining computationally efficient for potential edge-based deployment. This study benchmarks three Convolutional Neural Network (CNN) models for classification (VGG16, ResNet50, and EfficientNet-B0) and two encoder–decoder models for segmentation (U-Net and DeepLabV3+) using annotated pineapple and strawberry image datasets. A 5-fold cross-validation strategy was applied to ensure statistical robustness, with evaluation metrics including accuracy, precision, recall, F1-score, Intersection over Union (IoU), and Dice coefficient. Statistical significance was verified using the Friedman and Wilcoxon signed-rank tests (α = 0.05 and 0.01). EfficientNet-B0 achieved the best classification results with average accuracies of 91.4% (strawberry) and 90.7% (pineapple), significantly outperforming ResNet50 and VGG16 (p < 0.01). For segmentation, DeepLabV3+ obtained the highest performance with mean IoU values of 91.7% and 90.8% and Dice coefficients above 92%, indicating precise boundary delineation of ripe and defective regions. Computational efficiency analysis further showed that EfficientNet-B0 had the lowest inference time (0.026 s) and smallest model size (20.4 MB), making it ideal for real-time or embedded applications. Visual analysis confirmed that DeepLabV3+ maintained robustness at fruit boundaries, though minor misclassifications were observed. This benchmarking highlights the combination of EfficientNet-B0 and DeepLabV3+ as a reliable baseline for deep learning-based fruit quality assessment.