CycleGAN-Based Image Translation for Near-Infrared Camera-Trap Image Recognition

Renwu Gao, Siting Zheng, Jia He, Linlin Shen

Research output: Chapter in Book/Conference proceedingConference contributionpeer-review

1 Citation (Scopus)

Abstract

Due to its invisibility, NIR (Near-infrared) flash has been widely used to capture the images of wild animals in the night. Although the animals can be captured without notice, the gray NIR images are short of color and texture information and thus is difficult to analyze, for both human and machine. In this paper, we propose to use CycleGAN (Generative Adversarial Networks) to translate NIR image to the incandescent domain for visual quality enhancement. Example translations show that both color and texture can be well recovered by the proposed CycleGAN model. The recognition performance of a SSD based detector on the translated incandescent images is also significantly better than that on the original NIR images. Taking Wildebeest and Zebra for example, an increase of 16 % and 8 % in recognition accuracy has been observed.

Original languageEnglish
Title of host publicationPattern Recognition and Artificial Intelligence - International Conference, ICPRAI 2020, Proceedings
EditorsYue Lu, Nicole Vincent, Pong Chi Yuen, Wei-Shi Zheng, Farida Cheriet, Ching Y. Suen
PublisherSpringer Science and Business Media Deutschland GmbH
Pages453-464
Number of pages12
ISBN (Print)9783030598297
DOIs
Publication statusPublished - 2020
Externally publishedYes
Event2nd International Conference on Pattern Recognition and Artificial Intelligence, ICPRAI 2020 - Zhongshan, China
Duration: 19 Oct 202023 Oct 2020

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume12068 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference2nd International Conference on Pattern Recognition and Artificial Intelligence, ICPRAI 2020
Country/TerritoryChina
CityZhongshan
Period19/10/2023/10/20

Keywords

  • Deep learning
  • Generative adversarial networks
  • Image translation

ASJC Scopus subject areas

  • Theoretical Computer Science
  • General Computer Science

Fingerprint

Dive into the research topics of 'CycleGAN-Based Image Translation for Near-Infrared Camera-Trap Image Recognition'. Together they form a unique fingerprint.

Cite this