Scholarlink

<p>In image processing and computer vision, the fusion of infrared and visible images has emerged as a potent technique for enhancing segmentation accuracy and improving adversarial attack detection. Infrared and visible image fusion seeks to produce detailed outputs that highlight important objects and preserve fine details under challenging conditions. However, many current methods prioritize appearance and numerical evaluation while overlooking the requirements of advanced analytical tasks. To address these challenges, this research introduces Adversarial Ensemble Capsnet Equilibrium Graph neural Network (AECEGN) is designed to improve the system’s resilience and stability when exposed to malicious inputs or disruptions. The process begins with employing the Global context Encoder with Residual dense self-Attention Skill convoluted Network (GERASN) framework, which enhances the reliability of semantic segmentation models against adversarial attacks by leveraging both global and local information from infrared and visible images. The Adaptive Sand Cat Optimization (ASCO) technique is used for decomposition-based semantic segmentation in fused images to extract textural details and backgrounds, which optimizes the training objective like validation loss, normal training, and attacked training losses. The proposed method achieves a 96.53&#xa0;mIoU score in the MFNet dataset and achieves 97.39&#xa0;mIoU score in the M3FD dataset, respectively, which demonstrates superior performance than existing models.</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

An Efficient Fusion with Infrared and Visible Images for Deep Learning Based Adversarial Attack Detection

  • 期刊论文

  • 发表日期：

    2025年9月4日

查看全文
  • P. Murugeswari,
  • P. K. A. Chitra,
  • M. Ramkumar,
  • S. Sampath Kumar

摘要

In image processing and computer vision, the fusion of infrared and visible images has emerged as a potent technique for enhancing segmentation accuracy and improving adversarial attack detection. Infrared and visible image fusion seeks to produce detailed outputs that highlight important objects and preserve fine details under challenging conditions. However, many current methods prioritize appearance and numerical evaluation while overlooking the requirements of advanced analytical tasks. To address these challenges, this research introduces Adversarial Ensemble Capsnet Equilibrium Graph neural Network (AECEGN) is designed to improve the system’s resilience and stability when exposed to malicious inputs or disruptions. The process begins with employing the Global context Encoder with Residual dense self-Attention Skill convoluted Network (GERASN) framework, which enhances the reliability of semantic segmentation models against adversarial attacks by leveraging both global and local information from infrared and visible images. The Adaptive Sand Cat Optimization (ASCO) technique is used for decomposition-based semantic segmentation in fused images to extract textural details and backgrounds, which optimizes the training objective like validation loss, normal training, and attacked training losses. The proposed method achieves a 96.53 mIoU score in the MFNet dataset and achieves 97.39 mIoU score in the M3FD dataset, respectively, which demonstrates superior performance than existing models.