AgroQuali-FSL: a few-shot deep learning framework with QualiProtoNet for automated quality grading of fruits and vegetables.

As the supply chain moves towards uniformity, objectivity, and scalability, it is imperative that fruit and vegetable quality grading be conducted using automated methods. Traditional computer vision-based systems are effective only in controlled environments, but are commodity-specific and data-hungry; they require large annotated datasets to train for new defect types or emerging quality categories, making them less flexible. Most models based on CNNs and transformers are highly representational but do not generalise well in scenarios with few labelled samples, making them infeasible for real-life agricultural conditions where quality patterns change rapidly. These constraints further emphasise the urgent need for a flexible architecture to perform resilient quality grading under low-data conditions. In response, this work introduces AgroQuali-FSL, a hybrid deep learning and few-shot learning framework that incorporates an AgroVision-Backbone (a CNN-Transformer feature encoder) and a QualiProtoNet (a prototype-based classification module trained via episodic N-way K-shot learning). We propose a method that integrates a decision fusion scheme to merge the confidence scores from the backbones and the distance to the prototypes, thereby improving prediction stability, and also summarises the model's reasoning in an interpretable manner via Grad-CAM visualisations. The experiments on the Fruits Fresh and Rotten dataset and multiple defect datasets validate the effectiveness of our model, as it outperforms baseline CNN, ViT and ResNet models with substantial gains on both supervised and few-shot settings. AgroQuali-FSL, in particular, gains ∼3-7% accuracy in 1- and 5-shot cases and ∼2-4% gains via decision fusion-based refinement. Our results further demonstrate that AgroQuali-FSL is a resilient, scalable, and explainable system that can be tested in real-world agricultural grading systems and requires only a small amount of labelled data for fast adaptation to new types of commodities and defects.