Performance of Lightweight Neural Networks in Classification of Soy Bean Variety

Abstract

This research demonstrates the integration of image processing techniques with lightweight neural network models for the classification of Philippine soybean variants using morphological features. The canny edge detection algorithm was used to extract discriminative features, ensuring high-quality inputs for the classification phase. Among the evaluated neural network architectures, TabNet achieved the highest accuracy and balanced performance metrics (87.7%), benefiting from its attention-based mechanism for dynamic feature selection. However, considering computational efficiency and deployment constraints, the Shallow DNN model was identified as optimal due to its minimal model size (35.95 KB) and competitive performance metrics (85.9% accuracy). The NODE model showed acceptable performance but exhibited limitations due to higher false-negative rates, suggesting opportunities for future optimization. Overall, this study indicates that employing lightweight neural network models, particularly TabNet and Shallow DNN, provides a promising pathway for robust and efficient soybean classification, facilitating practical deployment in precision agriculture and resource-limited environments.