中文

基于颜色先验和视觉转换器的深度展开网络 UIE-UnFold

计算机视觉与模式识别 2024-08-21 v1

摘要

水下图像增强 (UIE) 在 various marine applications 中发挥着 crucial role,但由于 complex underwater environment,仍具有挑战性。当前的 learning-based approaches 常缺乏对 underwater image formation 中涉及的 physical processes 的 explicit incorporation,导致 optimization 受限,尽管 enhancement results 令人印象深刻。本文提出一种 novel deep unfolding network (DUN) 用于 UIE,该 network 将 color priors 和 inter-stage feature transformation 集成以提高 enhancement performance。 proposed DUN model 将 model-based methods 的 iterative optimization 和 reliability 与 deep learning 的 flexibility 和 representational power 结合起来,为 existing learning-based UIE approaches 提供了更 explainable 和 stable 的 solution。 proposed model 包含三个 key components: Color Prior Guidance Block (CPGB),其在 degraded 和 original images 的 color channels 之间建立 mapping;Nonlinear Activation Gradient Descent Module (NAGDM),其模拟 underwater image degradation process;以及 Inter Stage Feature Transformer (ISF-Former),其在 different network stages 之间 facilitate feature exchange。通过显式地 incorporating color priors 并建模 underwater image formation 的 physical characteristics,proposed DUN model 实现了 more accurate 和 reliable 的 enhancement results。在多个 underwater image 数据集上的 extensive experiments 显示,proposed model 在 quantitative 和 qualitative evaluations 中优于 state-of-the-art methods。proposed DUN-based approach 为 UIE 提供了有前景的 solution,使 marine research 中的 more accurate 和 reliable scientific analysis 成为可能。代码可在 https://github.com/CXH-Research/UIE-UnFold 获取。

关键词

引用

@article{arxiv.2408.10653,
  title  = {UIE-UnFold: Deep Unfolding Network with Color Priors and Vision Transformer for Underwater Image Enhancement},
  author = {Yingtie Lei and Jia Yu and Yihang Dong and Changwei Gong and Ziyang Zhou and Chi-Man Pun},
  journal= {arXiv preprint arXiv:2408.10653},
  year   = {2024}
}

备注

Accepted by DSAA CIVIL 2024