Deep Learning-Based Real-Time Classification of Regional Garlic Varieties of Karnataka for Food Quality Assurance

Abstract

Being a medicinal food, garlic (Allium sativum) is used for its antimicrobial, antioxidant, cardio protective and immune-modulatory properties, as well as for culinary use as a spice; its nutraceutical contents vary systematically among cultivars, and correct identification of regional garlic varieties is also important to ensure consistency in nutraceutical quality, food-industry standardisation, and preventing adulteration or mislabelling in the functional-food supply chain. At present, in Karnataka, regional cultivars like Agrifound White (G-41), Yamuna Safed (G-1), Bhima Omkar and Bhima Purple are differentiated manually by visual inspection, which is subjective, time-consuming and requires skilled personnel, and is likely to create quality-control risks in the downstream process under varying light conditions. To overcome this deficiency, in this study a garlic-variety classification system was developed and evaluated based on image processing and deep learning to classify garlic varieties in real-time. First, bulb images were pre-processed with OpenCV, localised in video or uploaded images using a YOLOv8-based object detection network, then classified into their regional variety or not using a VGG19-based CNN trained on clove-pattern, skin-texture and colour features and presented in a simple, user-friendly PyQt5 desktop application without any technical skills required. The overall accuracy of the system was 91% on the 5 class evaluation set, with the F1 scores of three out of four varieties being above 0.90. The proposed system could help maintain the nutraceutical quality and authenticity of garlic brought into food and medicinal supply chain in Karnataka by providing consistent and cost-effective variety verification.