Scholarlink

<p>Early and precise identification of rice leaf diseases is necessary to enhance crop productivity and contribute to sustainable agricultural activities in accordance with the Sustainable Development Goal 2. This paper suggests a hybrid deep learning model that features the advantages of the ResNet152V2 on local feature representation and Vision Transformer on global contextual representation. To reduce the issue of class imbalance in the dataset, a Generative Adversarial Network (GAN) has been used to generate synthetic samples for minority disease groups. The suggested structure is tested on a 16-class rice leaf disease dataset the construction of which is based on the combination of two publicly available Kaggle datasets. Experimental findings indicate that the initial hybrid model has a 92 percent classification accuracy and 74 percent minority class recall, and around 144.8 million parameters and 150&#xa0;min of training. Training time is also raised by adding the training of a discriminator and generator, thus GAN-based augmentation raises the model performance to 97 percent and 88 percent minority class recall, which is an improvement of 14 percentage points in detecting disease classes with low representation. The findings allow concluding that the samples obtained with the help of GAN actually increase the diversity of the dataset and make the model more capable of recognizing minority types of diseases. On the whole, the hybrid CNN-Transformer framework suggested can be discussed as a powerful and scalable intelligent plant disease detection method and has possible future applications in the areas of precision agriculture and automated systems of crop monitoring.</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

ViTRes: Vision Transformer and ResNet-Based FusionNet for Rice Leaf Disease Detection System

  • 期刊论文

  • 发表日期：

    2026年7月30日

查看全文
  • D. Hemanth Kumar,
  • P. Samundiswary

摘要

Early and precise identification of rice leaf diseases is necessary to enhance crop productivity and contribute to sustainable agricultural activities in accordance with the Sustainable Development Goal 2. This paper suggests a hybrid deep learning model that features the advantages of the ResNet152V2 on local feature representation and Vision Transformer on global contextual representation. To reduce the issue of class imbalance in the dataset, a Generative Adversarial Network (GAN) has been used to generate synthetic samples for minority disease groups. The suggested structure is tested on a 16-class rice leaf disease dataset the construction of which is based on the combination of two publicly available Kaggle datasets. Experimental findings indicate that the initial hybrid model has a 92 percent classification accuracy and 74 percent minority class recall, and around 144.8 million parameters and 150 min of training. Training time is also raised by adding the training of a discriminator and generator, thus GAN-based augmentation raises the model performance to 97 percent and 88 percent minority class recall, which is an improvement of 14 percentage points in detecting disease classes with low representation. The findings allow concluding that the samples obtained with the help of GAN actually increase the diversity of the dataset and make the model more capable of recognizing minority types of diseases. On the whole, the hybrid CNN-Transformer framework suggested can be discussed as a powerful and scalable intelligent plant disease detection method and has possible future applications in the areas of precision agriculture and automated systems of crop monitoring.