Adaptive weighted multi-discriminator CycleGAN for underwater image enhancement

Jaihyun Park, David K. Han, Hanseok Ko

Research output: Contribution to journalArticle

Abstract

In this paper, we propose a novel underwater image enhancement method. Typical deep learning models for underwater image enhancement are trained by paired synthetic dataset. Therefore, these models are mostly effective for synthetic image enhancement but less so for real-world images. In contrast, cycle-consistent generative adversarial networks (CycleGAN) can be trained with unpaired dataset. However, performance of the CycleGAN is highly dependent upon the dataset, thus it may generate unrealistic images with less content information than original images. A novel solution we propose here is by starting with a CycleGAN, we add a pair of discriminators to preserve contents of input image while enhancing the image. As a part of the solution, we introduce an adaptive weighting method for limiting losses of the two types of discriminators to balance their influence and stabilize the training procedure. Extensive experiments demonstrate that the proposed method significantly outperforms the state-of-the-art methods on real-world underwater images.

Original languageEnglish
Article number200
JournalJournal of Marine Science and Engineering
Volume7
Issue number7
DOIs
Publication statusPublished - 2019 Jan 1

Fingerprint

Discriminators
Image enhancement
image enhancement
learning
Experiments
method
experiment

Keywords

  • Generative adversarial networks
  • Image enhancement
  • Underwater

ASJC Scopus subject areas

  • Civil and Structural Engineering
  • Water Science and Technology
  • Ocean Engineering

Cite this

Adaptive weighted multi-discriminator CycleGAN for underwater image enhancement. / Park, Jaihyun; Han, David K.; Ko, Hanseok.

In: Journal of Marine Science and Engineering, Vol. 7, No. 7, 200, 01.01.2019.

Research output: Contribution to journalArticle

@article{d6cf2f775911498cb79fd53b9ee74e60,
title = "Adaptive weighted multi-discriminator CycleGAN for underwater image enhancement",
abstract = "In this paper, we propose a novel underwater image enhancement method. Typical deep learning models for underwater image enhancement are trained by paired synthetic dataset. Therefore, these models are mostly effective for synthetic image enhancement but less so for real-world images. In contrast, cycle-consistent generative adversarial networks (CycleGAN) can be trained with unpaired dataset. However, performance of the CycleGAN is highly dependent upon the dataset, thus it may generate unrealistic images with less content information than original images. A novel solution we propose here is by starting with a CycleGAN, we add a pair of discriminators to preserve contents of input image while enhancing the image. As a part of the solution, we introduce an adaptive weighting method for limiting losses of the two types of discriminators to balance their influence and stabilize the training procedure. Extensive experiments demonstrate that the proposed method significantly outperforms the state-of-the-art methods on real-world underwater images.",
keywords = "Generative adversarial networks, Image enhancement, Underwater",
author = "Jaihyun Park and Han, {David K.} and Hanseok Ko",
year = "2019",
month = "1",
day = "1",
doi = "10.3390/jmse7070200",
language = "English",
volume = "7",
journal = "Journal of Marine Science and Engineering",
issn = "2077-1312",
publisher = "MDPI AG",
number = "7",

}

TY - JOUR

T1 - Adaptive weighted multi-discriminator CycleGAN for underwater image enhancement

AU - Park, Jaihyun

AU - Han, David K.

AU - Ko, Hanseok

PY - 2019/1/1

Y1 - 2019/1/1

N2 - In this paper, we propose a novel underwater image enhancement method. Typical deep learning models for underwater image enhancement are trained by paired synthetic dataset. Therefore, these models are mostly effective for synthetic image enhancement but less so for real-world images. In contrast, cycle-consistent generative adversarial networks (CycleGAN) can be trained with unpaired dataset. However, performance of the CycleGAN is highly dependent upon the dataset, thus it may generate unrealistic images with less content information than original images. A novel solution we propose here is by starting with a CycleGAN, we add a pair of discriminators to preserve contents of input image while enhancing the image. As a part of the solution, we introduce an adaptive weighting method for limiting losses of the two types of discriminators to balance their influence and stabilize the training procedure. Extensive experiments demonstrate that the proposed method significantly outperforms the state-of-the-art methods on real-world underwater images.

AB - In this paper, we propose a novel underwater image enhancement method. Typical deep learning models for underwater image enhancement are trained by paired synthetic dataset. Therefore, these models are mostly effective for synthetic image enhancement but less so for real-world images. In contrast, cycle-consistent generative adversarial networks (CycleGAN) can be trained with unpaired dataset. However, performance of the CycleGAN is highly dependent upon the dataset, thus it may generate unrealistic images with less content information than original images. A novel solution we propose here is by starting with a CycleGAN, we add a pair of discriminators to preserve contents of input image while enhancing the image. As a part of the solution, we introduce an adaptive weighting method for limiting losses of the two types of discriminators to balance their influence and stabilize the training procedure. Extensive experiments demonstrate that the proposed method significantly outperforms the state-of-the-art methods on real-world underwater images.

KW - Generative adversarial networks

KW - Image enhancement

KW - Underwater

UR - http://www.scopus.com/inward/record.url?scp=85069177883&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85069177883&partnerID=8YFLogxK

U2 - 10.3390/jmse7070200

DO - 10.3390/jmse7070200

M3 - Article

AN - SCOPUS:85069177883

VL - 7

JO - Journal of Marine Science and Engineering

JF - Journal of Marine Science and Engineering

SN - 2077-1312

IS - 7

M1 - 200

ER -