Few-Shot Learning for Binary Classification in Red Chilli Powder

Detecting adulteration in red chilli powder is important for maintaining food quality and safety. One of the common adulterants that is added to red chilli powder is brick powder in global markets. This work evaluates the effective-ness of a few-shot learning approach to distinguish between pure and adulterated chilli powder. A small dataset comprising 30 images per class was used to simulate a low-data scenario, out of which 10 images were used for unseen data for testing purposes, which allowed to test how well these models perform under limited data conditions. Five pretrained networks -ResNet-18, MobileNetV3, EfficientNet-B0, DenseNet-121 and ConvNeXt-Tiny, were fine-tuned by keeping the feature extractor frozen and training only the final classification layer. Despite the small dataset, the models performed well, with DenseNet-121 and EfficientNet-B0 reaching validation accuracies of 95% and 90%. These results suggest that few-shot learning can be a practical and cost-effective method for creating quick, image-based tools to check for adulteration in spice powders.