Morphology-Aware Lightweight Deep Learning for Explainable Rice Variety Classification

Abstract

This research shows a lightweight convolutional neural network (CNN) framework for automatic rice variety classification. The proposed model is multi-branch modules of Inception style, res connection, channel – spatial attention mechanism to improve feature extraction without increasing the architecture compact and efficient. A rice image dataset of 5 varieties is used for training and testing the network, and the network model structure and Hyperparameters are improved by a sequence of tests. Results show that the approach we have proposed is able to achieve very high classification performance, indicating good prospects of practical application to automated rice quality control. To improve the trustworthiness more, explainable AI techniques, namely Grad-CAM, LIME and SHAP are being utilized, and the visualizations show that the model's predictions are primarily influenced by grain shape and texture features.