A novel residual learning of multi-scale feature extraction model for the classification of rice grain varieties

Rice serves as a fundamental food source for 50% of the world’s population highlighting its crucial role in ensuring food security. Deep learning has become crucial tools in automating the labor-intensive task of rice grain classification, using digital image processing to evaluate quality and grain variety. This work utilized a large dataset consisting of 75,000 images of five different rice grain varieties. There are 15,000 images for each type, which capture distinct texture, form, and color features. Image augmentation approaches, such as normalization and transformations, are utilized to enhance model robustness and mitigate overfitting. The study presented a novel ensemble model that combined a customized attention mechanism with modified residual learning and multi-scale feature learning of parallel filters networks to improve the ability of features extraction and classification of rice grain varieties. A wide range of performance criteria is employed to assess the effectiveness of the model. The ensemble model demonstrated outstanding competence in classification tasks, achieving accuracy values close to 99%. The Grad-CAM visualization validates the model’s attention towards pertinent characteristics among different rice grain varieties. The ensemble model outperformed pre-trained models and other works in terms of loss, accuracy, and F1-score, as shown by comparative analysis. This study enhances the field of agricultural informatics by boosting the accuracy of rice grain classification and food quality in general.