Scholarlink

<p>High intra-class variability, inadequate modelling of spatial dependencies and limitation of training samples results into various challenges while working with satellite images, to resolve these issues this research proposes DeepCAS (Deep Conditional Attentional Smoothing), a generative network that effectively utilises self-attention mechanisms, conditional generation, and label smoothing for satellite image synthesis. DeepCAS strategically positions the self-attention layers at ideal resolutions in conjunction with regularisation and conditional batch normalisation methods tailored to the properties of satellite images. Experiments conducted on EuroSAT dataset demonstrates the efficiency of DeepCAS across multiple evaluation parameters. The proposed architecture overcomes is issues of existing methods by enabling controllable generation of diverse, high-fidelity satellite images while maintaining training stability and class consistency.</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

DeepCAS: a novel GAN architecture for high-fidelity satellite image generation

  • 期刊论文

  • 发表日期：

    2025年11月21日

查看全文
  • Himani Deshpande,
  • Aryan Gurav,
  • Yathharth Karanjikar,
  • Nabanita Mandal

摘要

High intra-class variability, inadequate modelling of spatial dependencies and limitation of training samples results into various challenges while working with satellite images, to resolve these issues this research proposes DeepCAS (Deep Conditional Attentional Smoothing), a generative network that effectively utilises self-attention mechanisms, conditional generation, and label smoothing for satellite image synthesis. DeepCAS strategically positions the self-attention layers at ideal resolutions in conjunction with regularisation and conditional batch normalisation methods tailored to the properties of satellite images. Experiments conducted on EuroSAT dataset demonstrates the efficiency of DeepCAS across multiple evaluation parameters. The proposed architecture overcomes is issues of existing methods by enabling controllable generation of diverse, high-fidelity satellite images while maintaining training stability and class consistency.