Scholarlink

<p>In many real-world domains such as medical imaging, satellite vision, and rare object detection, acquiring large-scale annotated datasets is often expensive, time-consuming, or infeasible due to privacy and data scarcity constraints. This presents a significant challenge for training deep learning models that depend on diverse and representative data. This study aims to develop a generative model capable of synthesizing high-quality, structurally coherent, and diverse images in settings with limited data availability. The approach investigates whether integrating attention mechanisms and style modulation within a progressively growing GAN framework can enhance image realism and intra-class variation under constrained data conditions. An improved generative architecture, termed Style-Attentive Progressive Growing GAN (SAPGAN), is introduced. The model extends the conventional Progressive Growing GAN by incorporating two key modules: Adaptive Feature Fusion Attention (AFFA) for context-aware spatial and channel attention, and Adaptive Instance Normalization (AdaIN) for seman- tic style modulation across progressive resolution stages. These components are integrated within a progressively growing generator to facilitate stable training and multi-scale feature refinement. Experimental results indicate that SAPGAN consistently surpasses baseline models such as PGGAN and HieGAN across mul- tiple image resolutions. Quantitative evaluations using (Activation Maximization Score) AM Score and Mode Score demonstrate improvements in realism, fidelity, and diversity of the generated outputs. The architecture also provides enhanced semantic control and structural accuracy in synthesized images, particularly in medical imaging scenarios. These findings suggest that the combination of attention-based feature fusion and adaptive normalization significantly improves. generative performance in data-scarce environments. The proposed SAPGAN framework offers a promising direction for applications in data augmentation, syn- thetic dataset generation, and downstream visual learning tasks in low-resource domains.</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

SAPGAN: Style-Attentive Progressive Growing Generative Adversarial Network for Limited-Data Image Synthesis

  • 期刊论文

  • 发表日期：

    2026年3月3日

查看全文
  • Sweta Jha,
  • Namrata Govind Ambekar,
  • Surmila Thokchom

摘要

In many real-world domains such as medical imaging, satellite vision, and rare object detection, acquiring large-scale annotated datasets is often expensive, time-consuming, or infeasible due to privacy and data scarcity constraints. This presents a significant challenge for training deep learning models that depend on diverse and representative data. This study aims to develop a generative model capable of synthesizing high-quality, structurally coherent, and diverse images in settings with limited data availability. The approach investigates whether integrating attention mechanisms and style modulation within a progressively growing GAN framework can enhance image realism and intra-class variation under constrained data conditions. An improved generative architecture, termed Style-Attentive Progressive Growing GAN (SAPGAN), is introduced. The model extends the conventional Progressive Growing GAN by incorporating two key modules: Adaptive Feature Fusion Attention (AFFA) for context-aware spatial and channel attention, and Adaptive Instance Normalization (AdaIN) for seman- tic style modulation across progressive resolution stages. These components are integrated within a progressively growing generator to facilitate stable training and multi-scale feature refinement. Experimental results indicate that SAPGAN consistently surpasses baseline models such as PGGAN and HieGAN across mul- tiple image resolutions. Quantitative evaluations using (Activation Maximization Score) AM Score and Mode Score demonstrate improvements in realism, fidelity, and diversity of the generated outputs. The architecture also provides enhanced semantic control and structural accuracy in synthesized images, particularly in medical imaging scenarios. These findings suggest that the combination of attention-based feature fusion and adaptive normalization significantly improves. generative performance in data-scarce environments. The proposed SAPGAN framework offers a promising direction for applications in data augmentation, syn- thetic dataset generation, and downstream visual learning tasks in low-resource domains.