Scholarlink

<p>Accurate detection of potato leaf diseases remains challenging due to limited annotated data, class imbalance, and complex outdoor environmental changes caused by real-field conditions. In order to confront these problems, this work puts forward a comprehensive deep learning methodology that combines conditional generative adversarial networks (cGAN), reinforcement learning-based optimization, and an attention-enhanced CNN-LSTM structure. As a first step, the cGAN produces synthetic examples that are specific to each class to solve the problem of data imbalance and to expose the model to a higher diversity of features. Then, a deep Q-learning operator dynamically determines the essential parameters, such as the rate of learning and the size of the batch, to secure a stable and efficient conducting of the sessions of training. At last, the attention-enhanced CNN-LSTM system is able to learn complicated spatial and temporal features as well as to concentrate selectively on the regions of the leaves that are relevant to the disease. The framework put forward here has been tested on various sets of data, namely, the PlantVillage dataset, a six-class potato disease dataset, and a locally recorded real-field dataset. Results obtained through experimentation show that the method presented here is superior to the ones documented in the literature in terms of accuracy, robustness, and capacity for generalization, thereby attesting to its suitability for the very demanding deployment of agriculture in the real world. Moreover, the very fast speed of inference that the optimized design reaches makes it a very good option for the agricultural devices of low power at the edge.</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

A Unified cGAN-Reinforcement Learning-Based Optimized Attention CNN-LSTM Framework for Robust Potato Disease Detection and Classification

  • 期刊论文

  • 发表日期：

    2026年5月19日

查看全文
  • Ayush Kumar,
  • Naveen Kumar Tiwari,
  • Abhishek Bajpai

摘要

Accurate detection of potato leaf diseases remains challenging due to limited annotated data, class imbalance, and complex outdoor environmental changes caused by real-field conditions. In order to confront these problems, this work puts forward a comprehensive deep learning methodology that combines conditional generative adversarial networks (cGAN), reinforcement learning-based optimization, and an attention-enhanced CNN-LSTM structure. As a first step, the cGAN produces synthetic examples that are specific to each class to solve the problem of data imbalance and to expose the model to a higher diversity of features. Then, a deep Q-learning operator dynamically determines the essential parameters, such as the rate of learning and the size of the batch, to secure a stable and efficient conducting of the sessions of training. At last, the attention-enhanced CNN-LSTM system is able to learn complicated spatial and temporal features as well as to concentrate selectively on the regions of the leaves that are relevant to the disease. The framework put forward here has been tested on various sets of data, namely, the PlantVillage dataset, a six-class potato disease dataset, and a locally recorded real-field dataset. Results obtained through experimentation show that the method presented here is superior to the ones documented in the literature in terms of accuracy, robustness, and capacity for generalization, thereby attesting to its suitability for the very demanding deployment of agriculture in the real world. Moreover, the very fast speed of inference that the optimized design reaches makes it a very good option for the agricultural devices of low power at the edge.