Scholarlink

<p>Optical flow estimation, a fundamental task in computer vision, faces significant challenges in low-light environments due to issues such as low signal-to-noise ratios. Traditional methods rely on photometric consistency, which often fails in low-light conditions. Recent deep learning-based approaches, while effective under normal lighting, struggle with the noise and texture loss present in low-light images. In this paper, we propose a novel method that employs implicit guidance for enhancing low-light images to improve optical flow estimation. Our approach utilizes a channel-attention-based image enhancement network to reduce noise and improve image quality, followed by an iterative optical flow estimation module. During training, the enhancement network is supervised by features extracted from a pre-trained network on normal-light images, and the optical flow estimation loss is backpropagated to refine the enhancement process. Experiments on both synthetic and real low-light datasets demonstrate that our method significantly outperforms other methods, highlighting the effectiveness of implicit guidance in addressing the challenges of low-light optical flow estimation. Our code will be available at: <a href="https://github.com/HDU-ASL/IGEFlow">https://github.com/HDU-ASL/IGEFlow</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

Implicit guidance for enhancing low-light optical flow estimation via channel attention networks

  • 期刊论文

  • 发表日期：

    2025年2月27日

查看全文
  • Weichen Dai,
  • Hexing Wu,
  • Xiaoyang Weng,
  • Wanzeng Kong

摘要

Optical flow estimation, a fundamental task in computer vision, faces significant challenges in low-light environments due to issues such as low signal-to-noise ratios. Traditional methods rely on photometric consistency, which often fails in low-light conditions. Recent deep learning-based approaches, while effective under normal lighting, struggle with the noise and texture loss present in low-light images. In this paper, we propose a novel method that employs implicit guidance for enhancing low-light images to improve optical flow estimation. Our approach utilizes a channel-attention-based image enhancement network to reduce noise and improve image quality, followed by an iterative optical flow estimation module. During training, the enhancement network is supervised by features extracted from a pre-trained network on normal-light images, and the optical flow estimation loss is backpropagated to refine the enhancement process. Experiments on both synthetic and real low-light datasets demonstrate that our method significantly outperforms other methods, highlighting the effectiveness of implicit guidance in addressing the challenges of low-light optical flow estimation. Our code will be available at: https://github.com/HDU-ASL/IGEFlow.