Towards simulating foggy and hazy images and evaluating their authenticity

Ning Zhang, Lin Zhang*, Zaixi Cheng

*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingConference contributionAcademicpeer-review

23 Citations (Scopus)

Abstract

To train and evaluate fog/haze removal models, it is highly desired but burdensome to collect a large-scale dataset comprising well-aligned foggy/hazy images with their fog-free/haze-free versions. In this paper, we propose a framework, namely Foggy and Hazy Images Simulator (FoHIS for short), to simulate more realistic fog and haze effects at any elevation in images. What’s more, no former studies have introduced objective methods to evaluate the authenticity of synthetic foggy/hazy images. We innovatively design an Authenticity Evaluator for Synthetic foggy/hazy Images (AuthESI for short) to objectively measure which simulation algorithm could achieve more natural-looking results. We compare FoHIS with another two state-of-the-art methods, and the subjective results show that it outperforms those competitors. Besides, the prediction on simulated image’s authenticity made by AuthESI is highly consistent with subjective judgements (Source codes are publicly available at https://github.com/noahzn/FoHIS).
Original languageEnglish
Title of host publicationInternational Conference on Neural Information Processing
PublisherSpringer
Pages405-415
Number of pages11
DOIs
Publication statusPublished - 28 Oct 2017
Externally publishedYes
EventInternational Conference on Neural Information Processing, ICONIP 2017 -
Duration: 14 Sept 201718 Sept 2017

Publication series

NameLecture Notes in Computer Science
PublisherSpringer
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

ConferenceInternational Conference on Neural Information Processing, ICONIP 2017
Period14/09/1718/09/17

Keywords

  • ITC-CV

Fingerprint

Dive into the research topics of 'Towards simulating foggy and hazy images and evaluating their authenticity'. Together they form a unique fingerprint.

Cite this