ED-Net: Unified Enhancement-Denoising Deep Convolutional Network for Low-Light Mining Images

EasyChair Preprint 14661

Abstract

The mining environment poses unique challenges for image processing due to low illumination conditions and the presence of various types of noise. Existing image enhancement methods typically focus on either brightness enhancement or noise reduction, but rarely address both issues simultaneously. This paper proposes ED-Net, a unified model that simultaneously tackles both brightness enhancement and noise reduction through a novel joint loss function design. Addressing the lack of open-source mining image datasets, the authors introduce a method for constructing a low-illumination noise dataset to simulate the mining environment. Evaluated on the BSDS500 dataset, ED-Net demonstrates superior performance compared to state-of-the-art algorithms in terms of PSNR, SSIM, VIF metrics, and visual quality. The key innovations lie in the unified brightness enhancement and noise reduction model, the effective joint loss function, and the low-illumination noise dataset construction method for simulating mining images.

Keyphrases: Deep convolution, Image denoising, Low-light image enhancement, Mine Image Enhancement, brightness enhancement and noise reduction, histogram matching, lack of open source, loss function, underground mining images

@booklet{EasyChair:14661,
  author    = {Jiaqi Wu and Da Lu and Yu Tao and Hui Ding and Guoping Huo},
  title     = {ED-Net: Unified Enhancement-Denoising Deep Convolutional Network for Low-Light Mining Images},
  howpublished = {EasyChair Preprint 14661},
  year      = {EasyChair, 2024}}