Scholarlink

Bacterial Leaf Blight, Brown Spot, Rice Blast, Leaf Scald, and Narrow Brown Spot are major rice diseases that significantly reduce yield and grain quality. Early and accurate diagnosis is essential for effective disease management. In this study, five deep learning models—InceptionV3, MobileNetV3Small, MobileNetV3Large, EfficientNetB0, and ResNet50—were individually trained and evaluated. Attention mechanisms, including the Convolutional Block Attention Module (CBAM) and Squeeze-and-Excitation (SE), were subsequently integrated into the backbones to enhance classification performance. The two best-performing models, ResNet50 and EfficientNetB0, were then combined to construct a hybrid architecture, REBANet (ResNet–EfficientNet–Attention Network). REBANet leverages the complementary representational strengths of both backbones and employs attention modules to refine channel- and spatial-wise feature selection, enabling the model to focus on disease-salient regions. Experimental results show that REBANet achieves 99.62% accuracy, outperforming all standalone backbones. In addition, Grad-CAM is used to visualize and interpret the model’s diagnostic decisions.

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

REBANet: A Hybrid Deep Learning Model for Rice Leaf Disease Prediction with Grad-CAM Explainability

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Nguyen Minh Khiem,
  • Huynh Trong The,
  • Pham Ngoc Quyen

摘要

Bacterial Leaf Blight, Brown Spot, Rice Blast, Leaf Scald, and Narrow Brown Spot are major rice diseases that significantly reduce yield and grain quality. Early and accurate diagnosis is essential for effective disease management. In this study, five deep learning models—InceptionV3, MobileNetV3Small, MobileNetV3Large, EfficientNetB0, and ResNet50—were individually trained and evaluated. Attention mechanisms, including the Convolutional Block Attention Module (CBAM) and Squeeze-and-Excitation (SE), were subsequently integrated into the backbones to enhance classification performance. The two best-performing models, ResNet50 and EfficientNetB0, were then combined to construct a hybrid architecture, REBANet (ResNet–EfficientNet–Attention Network). REBANet leverages the complementary representational strengths of both backbones and employs attention modules to refine channel- and spatial-wise feature selection, enabling the model to focus on disease-salient regions. Experimental results show that REBANet achieves 99.62% accuracy, outperforming all standalone backbones. In addition, Grad-CAM is used to visualize and interpret the model’s diagnostic decisions.