Dual-Branch Shape Texture Learning With Pyramid Residual and Sobel Edge Layers for Rice Variety Classification

Abstract

Automated rice variety classification categorizes rice types based on features, such as grain texture, shape, color, and size. Traditional handcrafted feature-based models fail at rice classification due to lighting variations, background noise, and environmental conditions, leading to overlapping visual characteristics and frequent misclassification. To address these challenges, this research proposes a novel Dual-Branch Convolutional Neural Network with Pyramid Residual Units and a Sobel Edge Layer (DB-PRU-SEL) for rice variety classification. The DB-PRU-SEL consists of two branches: The first extracts robust shape features using hierarchical PRU to capture multi-scale structural details, while the second focuses on texture by integrating Sobel edge detection with convolutional layers. The outputs from both branches are concatenated by an attention-guided feature fusion mechanism that selectively enhances the discriminative features before passing them through fully connected layers for final classification. DB-PRU-SEL achieves superior performance on a rice image dataset, attaining an accuracy of 99.87% and an Area Under the Curve (AUC) of 99.99%, respectively. Comparative and ablation studies demonstrate that DB-PRU-SEL outperforms conventional classifiers. Additionally, complexity analysis indicates reduced training time and memory usage, making it an efficient solution for rice variety classification.