Scholarlink

<p>Shadows degrade image quality by obscuring textures, distorting colors, and reducing luminance. Existing methods often struggle to fully restore shadowed regions due to their reliance on basic illumination enhancement. This paper introduces VAE-mSRNet, a Collaborative Feature Disentanglement network that extends the multi-scale shadow removal framework mSRNet and decomposes shadowed and shadow-free images into shared and distinct feature components. By integrating a Variational Autoencoder (VAE), VAE-mSRNet compactifies the feature space and strengthens the separation of illumination-invariant textures and color cues. This results in a more structured representation that improves both separation quality and feature reconstruction. Our approach addresses challenges in shadow removal, such as texture inconsistency and color imbalance. Comprehensive experiments on benchmark datasets (AISTD and SRD) demonstrate that VAE-mSRNet achieves competitive and in many cases leading results across PSNR and RMSE metrics, with particular strength in shadow-region restoration, while maintaining fine texture details and color fidelity in complex shadow scenarios. Here, we show that VAE-mSRNet offers a robust solution for shadow removal, setting a new benchmark in the field. The source code and pre-trained models are available at <a href="https://github.com/Midrees330/VAE_MSRNet">https://github.com/Midrees330/VAE_MSRNet</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

Latent-optimized collaborative feature disentanglement for enhanced shadow removal

  • 期刊论文

  • 发表日期：

    2026年7月4日

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

摘要

Shadows degrade image quality by obscuring textures, distorting colors, and reducing luminance. Existing methods often struggle to fully restore shadowed regions due to their reliance on basic illumination enhancement. This paper introduces VAE-mSRNet, a Collaborative Feature Disentanglement network that extends the multi-scale shadow removal framework mSRNet and decomposes shadowed and shadow-free images into shared and distinct feature components. By integrating a Variational Autoencoder (VAE), VAE-mSRNet compactifies the feature space and strengthens the separation of illumination-invariant textures and color cues. This results in a more structured representation that improves both separation quality and feature reconstruction. Our approach addresses challenges in shadow removal, such as texture inconsistency and color imbalance. Comprehensive experiments on benchmark datasets (AISTD and SRD) demonstrate that VAE-mSRNet achieves competitive and in many cases leading results across PSNR and RMSE metrics, with particular strength in shadow-region restoration, while maintaining fine texture details and color fidelity in complex shadow scenarios. Here, we show that VAE-mSRNet offers a robust solution for shadow removal, setting a new benchmark in the field. The source code and pre-trained models are available at https://github.com/Midrees330/VAE_MSRNet.