Scholarlink

<p>Image restoration tasks, such as shadow removal, deraining, and denoising, traditionally necessitate specialized architectures. This paper introduces a task-agnostic image restoration framework based on variational feature disentanglement, utilizing dual variational autoencoder (VAE) encoders to separate degradation-specific features from content-preserving representations. Our approach, integrating a dual-scale attention mechanism, degrade-aware decoder, and a degrade area localizer, demonstrates competitive performance across shadow removal (AISTD, SRD), deraining (Rain100L), and denoising (BSD68) tasks, achieving computational efficiency (58.39 GFLOPs, 79.64 FPS). Here, we show that our architecture, trained independently per task, effectively disentangles features without task-specific architectural components at inference, offering a scalable solution for diverse restoration scenarios. The source code and pre-trained models are available at <a href="https://github.com/Midrees330/VDTA_IRNet">https://github.com/Midrees330/VDTA_IRNet</a>.</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

Variational disentanglement for task-agnostic image restoration

  • 期刊论文

  • 发表日期：

    2026年4月15日

查看全文
  • Muhammad Idrees,
  • Ying Huang,
  • Ailin Li,
  • Jiahao Jin

摘要

Image restoration tasks, such as shadow removal, deraining, and denoising, traditionally necessitate specialized architectures. This paper introduces a task-agnostic image restoration framework based on variational feature disentanglement, utilizing dual variational autoencoder (VAE) encoders to separate degradation-specific features from content-preserving representations. Our approach, integrating a dual-scale attention mechanism, degrade-aware decoder, and a degrade area localizer, demonstrates competitive performance across shadow removal (AISTD, SRD), deraining (Rain100L), and denoising (BSD68) tasks, achieving computational efficiency (58.39 GFLOPs, 79.64 FPS). Here, we show that our architecture, trained independently per task, effectively disentangles features without task-specific architectural components at inference, offering a scalable solution for diverse restoration scenarios. The source code and pre-trained models are available at https://github.com/Midrees330/VDTA_IRNet.