Scholarlink

<p>This paper presents a deep neural network-based method to improve the visual quality of images captured under challenging lighting conditions, ranging from dark to over-lit scenarios, without paired samples. To simplify the deep neural network architecture without loss its performance, we focus on data preprocessing and employ a no-reference multilateral loss functions. The proposed loss function is designed to adjust overall brightness, preserve color ratio, and control region-based process by the exposure levels. Experimental results demonstrate that the proposed method is faster than state-of-the-art approaches, preserves color balance, prevents noise amplification in dark regions, and reveals details in overexposed areas without color artifacts. Furthermore, the method can handle a wide range of image brightness levels with lower computational complexity, making it suitable for online applications. The source code and links to the datasets are available at <a href="https://github.com/AlirezaKhajehvandi/EIB-FNDL">https://github.com/AlirezaKhajehvandi/EIB-FNDL</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

Enhancing image brightness without paired data: a fast no-reference deep learning method

  • 期刊论文

  • 发表日期：

    2025年4月12日

查看全文
  • Alireza Khajehvandi,
  • Mehdi Ezoji

摘要

This paper presents a deep neural network-based method to improve the visual quality of images captured under challenging lighting conditions, ranging from dark to over-lit scenarios, without paired samples. To simplify the deep neural network architecture without loss its performance, we focus on data preprocessing and employ a no-reference multilateral loss functions. The proposed loss function is designed to adjust overall brightness, preserve color ratio, and control region-based process by the exposure levels. Experimental results demonstrate that the proposed method is faster than state-of-the-art approaches, preserves color balance, prevents noise amplification in dark regions, and reveals details in overexposed areas without color artifacts. Furthermore, the method can handle a wide range of image brightness levels with lower computational complexity, making it suitable for online applications. The source code and links to the datasets are available at https://github.com/AlirezaKhajehvandi/EIB-FNDL.