Scholarlink

<p>Removing reflection from images taken through glass is widely used in various computer vision tasks. It is a highly ill-posed problem that requires an accurate focus on important features. The previous methods mostly use a fixed spatial receptive field and cannot extract the reflection features of images well. In order to obtain more accurate spatial features, we propose an enhanced network architecture that introduces multi-scale spatial attention mechanism. In addition, considering the complementary relationship between the transmission layer and the reflection layer, many works adopt the feature shunting method. However, discarded features are not fully applicable to another branch, which weakens the robustness of the overall model. To solve this problem, we improve the feature shunting module by extracting beneficial items in the feature interaction. We also design a new gradient mutual exclusion function to restore weak texture regions. Comparative analysis with state-of-the-art methods underscores the effectiveness of the proposed algorithm, particularly in preserving image details and maintaining visual quality. Our code is publicly available at <a href="https://github.com/knight936/Single-image-reflection-removal">https://github.com/knight936/Single-image-reflection-removal</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

Multi-scale spatial attention and network enhancement for single-image reflection removal

  • 期刊论文

  • 发表日期：

    2025年6月18日

查看全文
  • Tao Chen,
  • Qiliang Yang,
  • Yin Chen,
  • Qizhen Zhou

摘要

Removing reflection from images taken through glass is widely used in various computer vision tasks. It is a highly ill-posed problem that requires an accurate focus on important features. The previous methods mostly use a fixed spatial receptive field and cannot extract the reflection features of images well. In order to obtain more accurate spatial features, we propose an enhanced network architecture that introduces multi-scale spatial attention mechanism. In addition, considering the complementary relationship between the transmission layer and the reflection layer, many works adopt the feature shunting method. However, discarded features are not fully applicable to another branch, which weakens the robustness of the overall model. To solve this problem, we improve the feature shunting module by extracting beneficial items in the feature interaction. We also design a new gradient mutual exclusion function to restore weak texture regions. Comparative analysis with state-of-the-art methods underscores the effectiveness of the proposed algorithm, particularly in preserving image details and maintaining visual quality. Our code is publicly available at https://github.com/knight936/Single-image-reflection-removal.