Scholarlink

<p>India stands top in the world as a huge chilli exporter, yet crop yields are significantly impacted by bacterial, fungal, viral and environmental diseases. Accurate disease detection stands challenging due to limited data, complex backgrounds, poor image quality and multi-disease occurrence. While deep convolutional neural networks (DCNNs) promises that they suffer from overfitting, local optima convergence, and suboptimal hyperparameter selection. This article proposes a multi-class chilli leaf disease detection and classification framework combining conditional Variational Autoencoders (cVAEs) and a Swin Transformer based LSTM Classifier. By using a publicly available Kaggle dataset, images undergo pre-processing that include scaling, normalization, histogram equalization, noise removal and Albumentation-based augmentation which addresses class imbalance. In the pre-processing stage, however, the cVAE serves as a noise-removal module. The dataset has been splitted into 70% of training data, 20% of validation data along with 10% of testing data. Features extracted via Swin Transformer are fed into an LSTM classifier. The proposed model achieves 99.49% accuracy with augmentation and 97.56% without outperforming existing methods. Comparative analysis demonstrates superior classification performance, providing an effective diagnostic tool for phytopathologists and agricultural researchers.</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

Multi Class Classification of Chilli Leaf Diseases Through Conditional Variational Encoder (cVAE) and a Swin Transformer Based LSTM Classifier

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月12日

查看全文
  • Shiva Shankar Annaram,
  • G. Gopichand

摘要

India stands top in the world as a huge chilli exporter, yet crop yields are significantly impacted by bacterial, fungal, viral and environmental diseases. Accurate disease detection stands challenging due to limited data, complex backgrounds, poor image quality and multi-disease occurrence. While deep convolutional neural networks (DCNNs) promises that they suffer from overfitting, local optima convergence, and suboptimal hyperparameter selection. This article proposes a multi-class chilli leaf disease detection and classification framework combining conditional Variational Autoencoders (cVAEs) and a Swin Transformer based LSTM Classifier. By using a publicly available Kaggle dataset, images undergo pre-processing that include scaling, normalization, histogram equalization, noise removal and Albumentation-based augmentation which addresses class imbalance. In the pre-processing stage, however, the cVAE serves as a noise-removal module. The dataset has been splitted into 70% of training data, 20% of validation data along with 10% of testing data. Features extracted via Swin Transformer are fed into an LSTM classifier. The proposed model achieves 99.49% accuracy with augmentation and 97.56% without outperforming existing methods. Comparative analysis demonstrates superior classification performance, providing an effective diagnostic tool for phytopathologists and agricultural researchers.