Rice Variety Identification Based on Transfer Learning Architecture Using DENS-INCEP

Abstract:

Accurate identification of rice varieties is critical for agricultural quality control, food traceability, and crop management. Traditional manual classification methods are time-consuming, inconsistent, and prone to human error, necessitating the development of automated and reliable identification systems. This study proposes a hybrid deep learning architecture termed DENS-INCEP, which integrates the feature extraction capabilities of the pre-trained DenseNet-201 network with a custom Inception module to achieve high-accuracy rice variety classification. The DenseNet-201 backbone, pre-trained on ImageNet and frozen during training, serves as a robust feature extractor, capturing rich hierarchical visual representations from rice grain images. A multi-scale Inception module with four parallel convolutional branches (1×1, 3×3, 5×5, and MaxPool projections) is appended after the third dense block transition, enabling simultaneous extraction of fine-grained and coarse-grained textural features. The concatenated Inception features are subsequently processed through a Global Average Pooling layer, a fully connected layer with 502 neurons, and a final softmax classification head. Experiments were conducted on the Rice Image Dataset comprising five distinct varieties. The proposed DENS-INCEP architecture demonstrates superior classification performance, validating the effectiveness of combining dense connectivity with multi-scale feature fusion for agricultural image recognition tasks.

Keywords: Rice variety classification, transfer learning, DenseNet-201, Inception module, deep learning