Scholarlink

This study presents a comparative analysis of deep learning architectures for automated rice variety classification, focusing on whether Convolutional Neural Networks (CNNs) or transformer-based models offer superior performance. We evaluate three representative CNNs (MobileNetV2, ResNet18, and DenseNet121), a hybrid CNN model (ConvNeXt-Base), and two transformer variants (ViT-Base and Swin-Tiny). Experiments on the PaddyVarietyBD dataset, comprising 14,000 rice grain images across 35 varieties, assessed model performance using accuracy, precision, recall, and F1-score. Additionally, we apply Grad-CAM to visualize and interpret each model’s focus regions during classification. Results consistently show CNN-based models outperform their transformer counterparts in both classification accuracy and computational efficiency. DenseNet121 achieved the highest accuracy (97.32%), while MobileNetV2 provided competitive performance (96.79%) with substantially lower computational cost. Grad-CAM visualizations revealed that CNNs attend more coherently to key grain features, whereas transformer models, particularly ViT-Base, displayed more scattered attention patterns. These findings underscore the practical advantages of CNN architectures, especially lightweight models like MobileNetV2, for effective and interpretable rice variety classification.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

CNNs Versus Transformers: An Interpretability-Focused Study for Rice Variety Classification

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Phi-Hung Hoang,
  • Nam-Thuan Trinh,
  • Van-Manh Tran,
  • Thi-Thu-Hong Phan

摘要

This study presents a comparative analysis of deep learning architectures for automated rice variety classification, focusing on whether Convolutional Neural Networks (CNNs) or transformer-based models offer superior performance. We evaluate three representative CNNs (MobileNetV2, ResNet18, and DenseNet121), a hybrid CNN model (ConvNeXt-Base), and two transformer variants (ViT-Base and Swin-Tiny). Experiments on the PaddyVarietyBD dataset, comprising 14,000 rice grain images across 35 varieties, assessed model performance using accuracy, precision, recall, and F1-score. Additionally, we apply Grad-CAM to visualize and interpret each model’s focus regions during classification. Results consistently show CNN-based models outperform their transformer counterparts in both classification accuracy and computational efficiency. DenseNet121 achieved the highest accuracy (97.32%), while MobileNetV2 provided competitive performance (96.79%) with substantially lower computational cost. Grad-CAM visualizations revealed that CNNs attend more coherently to key grain features, whereas transformer models, particularly ViT-Base, displayed more scattered attention patterns. These findings underscore the practical advantages of CNN architectures, especially lightweight models like MobileNetV2, for effective and interpretable rice variety classification.