Scholarlink

<p>Fog is one of the most dangerous weather phenomena impacting road visibility and greatly elevating the chances of traffic incidents. For intelligent transportation systems and driver-assistant technologies, prompt and accurate determination of fog intensity is critical. This paper introduces a new model called AdaFogNet which combines images captured by VGG16 network with adaptive filters to classify them into three levels - No Fog, Medium Fog, Dense Fog. The Adaptive Fog Filter improves relevant features using Gaussian blur alongside Sobel edge detection and contrast enhancement applied with intensity-aware weighted averaging. Captured features processed through frozen VGG16 convolutional layers followed by custom multi-class dense layers designed for final classification. Experimental results further confirm the outstanding performance of AdaFogNet over state-of-the-art models including VGG16, InceptionNet, MobileNet V3 and ResNet50 achieving 97.73% in accuracy and 98.97% precision giving these systems diverse operational range under varying fog conditions fitting for real-world autonomous and surveillance system deployment. As fog is considered to be a dense aerosol layer of water droplets, the proposed AdaFogNet technique is a significant contribution to aerosol study by presenting a high accuracy classification framework for prediction of fog level in the atmosphere.</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

AdaFogNet: An Adaptive Preprocessing-Based Deep Learning Model for Multi-Class Fog Severity Classification

  • 期刊论文

  • 发表日期：

    2025年10月9日

查看全文
  • M. Muthulakshmi,
  • R. Jansi,
  • Tushar Kumar

摘要

Fog is one of the most dangerous weather phenomena impacting road visibility and greatly elevating the chances of traffic incidents. For intelligent transportation systems and driver-assistant technologies, prompt and accurate determination of fog intensity is critical. This paper introduces a new model called AdaFogNet which combines images captured by VGG16 network with adaptive filters to classify them into three levels - No Fog, Medium Fog, Dense Fog. The Adaptive Fog Filter improves relevant features using Gaussian blur alongside Sobel edge detection and contrast enhancement applied with intensity-aware weighted averaging. Captured features processed through frozen VGG16 convolutional layers followed by custom multi-class dense layers designed for final classification. Experimental results further confirm the outstanding performance of AdaFogNet over state-of-the-art models including VGG16, InceptionNet, MobileNet V3 and ResNet50 achieving 97.73% in accuracy and 98.97% precision giving these systems diverse operational range under varying fog conditions fitting for real-world autonomous and surveillance system deployment. As fog is considered to be a dense aerosol layer of water droplets, the proposed AdaFogNet technique is a significant contribution to aerosol study by presenting a high accuracy classification framework for prediction of fog level in the atmosphere.