Scholarlink

<p>In this paper, we present a new hybrid deep learning framework for the mustard leaf disease detection based on the Co-scale Convolutional attentional Transformers (CoaT), Capsule Networks (CapsNet) and Ensemble Stack learning (ESTACK). CoaT harnesses the power of each component: it captures multi-scale global and local semantic features by hierarchical attention; CapsNet provides spatial equivariance and pose awareness which is important for the recognition of disease with different orientations; and ESTACK gives a lightweight and non-iterative classification layer that greatly reduces computational time. We experimentally validated our work on a self-created Healthy-Disease Mustard Leaf Set (HDMLS) comprising of five classes, each comprising of several stages of white rust, Alternaria blight and healthy samples. Proper pre-processing and augmentation made the model very robust for the environmental variance. It is shown that the proposed model yielded 98.38% accuracy which is higher than the existing approaches namely MPNet: 97.11%, CNN_LSTM: 96.48%, ResNet50: 95.17%. Along with the accuracy, the model attained high performance with precision (98.43%), recall (98.39%), F1-score (98.41%) and AUC-ROC (98.52%).</p>

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

CoaT-CapsNet-ESTACK: a hybrid deep learning architecture for high-precision mustard leaf disease detection using multiscale attention and pose-aware feature encoding

  • 期刊论文

  • 发表日期：

    2025年12月8日

查看全文
  • Shilpa Juneja,
  • Parneeta Dhaliwal,
  • Atul Srivastava

摘要

In this paper, we present a new hybrid deep learning framework for the mustard leaf disease detection based on the Co-scale Convolutional attentional Transformers (CoaT), Capsule Networks (CapsNet) and Ensemble Stack learning (ESTACK). CoaT harnesses the power of each component: it captures multi-scale global and local semantic features by hierarchical attention; CapsNet provides spatial equivariance and pose awareness which is important for the recognition of disease with different orientations; and ESTACK gives a lightweight and non-iterative classification layer that greatly reduces computational time. We experimentally validated our work on a self-created Healthy-Disease Mustard Leaf Set (HDMLS) comprising of five classes, each comprising of several stages of white rust, Alternaria blight and healthy samples. Proper pre-processing and augmentation made the model very robust for the environmental variance. It is shown that the proposed model yielded 98.38% accuracy which is higher than the existing approaches namely MPNet: 97.11%, CNN_LSTM: 96.48%, ResNet50: 95.17%. Along with the accuracy, the model attained high performance with precision (98.43%), recall (98.39%), F1-score (98.41%) and AUC-ROC (98.52%).