A hybrid machine vision and handcrafted features fusion based approach for fine-grained millet classification

Millets are increasingly recognized as climate-resilient, nutrient-rich grains essential for food and nutritional security. However, the limited availability of annotated datasets and the visual similarities among millet varieties such as Barnyard Millet, Little Millet, and Proso Millet pose significant challenges in their accurate identification and classification. To address these issues, this study proposes two novel hybrid frameworks that leverage both handcrafted and deep learning paradigms. A key component shared by both frameworks is the incorporation of a custom-designed handcrafted feature extraction process. The first framework introduces a Stacked Ensemble Machine Learning approach, that integrates handcrafted techniques and machine learning classifiers as Support Vector Machine (SVM), Random Forest (RF), and K-Nearest Neighbor (KNN) classifiers using a multinomial logistic regression meta-learner. The second framework adopts a bidirectional deep learning architecture; wherein handcrafted feature vectors are fused with the pretrained VGG19 model to achieve effective classification of different millet varieties. The stacked ensemble hybrid model achieved the 97% classification accuracy where as the second model achieved the highest accuracy of 96.7%, an F1-score of 0.967, and a ROC-AUC of 0.995, while maintaining excellent tradeoff between training and inference times.