Multi-Level Hierarchical Approach for Banana Variety Classification and Grading at Bunch Level

Accurate classification and grading of banana varieties ensures quality standards in large-scale food processing. Traditional methods focused on hand or finger-level classification are insufficient for bunch-level grading, particularly for visually similar varieties. To address this gap, we introduce a novel multi-level hierarchical classification approach that utilizes transfer learning and advanced deep learning architectures such as MobileNet and EfficientNet. Our approach outperforms the conventional state-of-the-art (SOTA) single-shot multi-class classification methods, increasing accuracy from 94.88% to 98.66% by effectively handling the high inter-class similarity. This approach fills the gap in existing datasets and has significant implications for large-scale food processing and supply chains. Our novel classification system employs our unique bunch-level dataset comprising 25,909 images, the largest of its kind, to create more accurate and robust models, outperforming existing hand-level and finger-level classification.