Scholarlink

In this work, we offered ideas for an effective and generalizable framework, the mixed precision-based EfficientNet-B0 (MPEfficientNet-B0) model, to classify rice leaf diseases using the mixed precision technique in the base model EfficientNet-B0. The mixed precision training helps boost computational speed while conserving resources. Moreover, the proposed framework uses focal loss to tackle the issue of class disparity, leading to enhanced predictive performance. The dataset used to test the efficacy of this model is the rice leaf disease. The experimental study consists of state-of-the-art models such as XceptionNet, VGG16, ResNet50, InceptionV3, CNN-Based, SVM, Inception-ResNet-V2, ADSNN-BO, EfficientNet-B0. The proposed framework shows the best testing accuracy of 98.86% and low computational cost (measured in terms of Floating Point Operations per Second (FLOPs)) of 27.02M FLOPs. Furthermore, the heatmaps are used to present the attention map using Grad-CAM (Gradient-weighted Class Activation Mapping) showcasing how a model makes decisions, and which features of input are important while prediction.

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

An Optimal Framework for Rice Leaf Disease Prediction Using Mixed Precision and Gradient-Weighted Class Activation Mapping

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Preeti Raj Verma,
  • Deepika Pantola,
  • Navneet Pratap Singh

摘要

In this work, we offered ideas for an effective and generalizable framework, the mixed precision-based EfficientNet-B0 (MPEfficientNet-B0) model, to classify rice leaf diseases using the mixed precision technique in the base model EfficientNet-B0. The mixed precision training helps boost computational speed while conserving resources. Moreover, the proposed framework uses focal loss to tackle the issue of class disparity, leading to enhanced predictive performance. The dataset used to test the efficacy of this model is the rice leaf disease. The experimental study consists of state-of-the-art models such as XceptionNet, VGG16, ResNet50, InceptionV3, CNN-Based, SVM, Inception-ResNet-V2, ADSNN-BO, EfficientNet-B0. The proposed framework shows the best testing accuracy of 98.86% and low computational cost (measured in terms of Floating Point Operations per Second (FLOPs)) of 27.02M FLOPs. Furthermore, the heatmaps are used to present the attention map using Grad-CAM (Gradient-weighted Class Activation Mapping) showcasing how a model makes decisions, and which features of input are important while prediction.