Progressive feature enrichment and ensemble learning for enhanced rice seed purity classification

This study introduces an effective method for classifying rice seed purity varieties, addressing the challenge of identifying mixed seeds in real-world scenarios. Initially, conventional feature extraction methods such as Gray Level Co-occurrence Matrix (GLCM), Scale-Invariant Feature Transform (SIFT), Histogram of Oriented Gradients (HOG), and GIST are evaluated. However, these methods prove insufficient due to their inability to fully capture the complex characteristics of rice seeds. To address this, we propose a refined feature engineering strategy, optimizing domain-specific morphological, color, and textural features of rice seeds. Through an iterative refinement process, the feature set is systematically expanded. Initially, it comprises 18 basic features. Subsequently, we integrate color features, increasing the total to 22. Continued optimization with GLCM-based texture features leads to an expansion to 36 features, and finally, the set reaches 52 features, encompassing advanced morphological, color, and texture attributes. Notably, we also employ ensemble methods, particularly stacking, which combines the predictions of multiple models to create a final prediction. Experimental results demonstrate a significant improvement in classification accuracy compared to conventional extraction methods. Specifically, stacking boosts accuracy to 97.58%. These findings underscore the superiority of customized feature design over generic extractors, providing a robust solution for practical rice seed purity identification.