Few-Shot Learning for Strawberry Variety Classification: A Prototypical Network-Based Approach

Classification of different types of strawberries is an essential problem in agriculture and quality control; however, it presents a challenge to traditional deep learning solutions because of the need for large-scale labeled datasets. In this research, a few-shot learning (FSL) model based on Prototypical Networks was applied to classify four different commercial types of strawberries (A, B, C, and Jumbo), utilizing the Mendeley Strawberry Grading Dataset that includes only 103 labeled images for training. A unique pipeline for training an FSL model was created through a combination of warm-up pre-training, differential learning rate strategy, and 4-way 5-shot episodic training; two backbones (ResNet50 and EfficientNet-B0) were compared under a 5-fold stratified cross-validation. While ResNet50 obtained test accuracy, macro F1-score, and Cohen's Kappa equal to 99.00%, 99.11%, and 98.62% respectively, EfficientNet-B0 achieved 98.00%, 98.33%, and 97.30%. Both models produced an AUC value of 99.80%. Interpretability analysis utilizing GradCAM++ showed that both networks focused on relevant areas of the fruits. This research proved that metric-based few-shot learning can be used successfully as an alternative solution to traditional deep learning methods in case of scarce datasets of agricultural images.