Depth estimation for hazy images using deep learning

Laksmita Rahadianti, Fumihiko Sakaue, Jun Sato

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

Abstract

3D scene understanding is important for many applications in the computer vision field. However, the majority of existing solutions commonly assume the images to be captured in clear media. In real world situations, we may encounter less than ideal conditions, for example haze or fog. In these cases, the captured images will contain scattering and veiling effects that obscure the features of the scene. Many studies approach these images by first removing the scattering effects to obtain an approximate clear image. However, by studying the physical model of light propagation in scattering media, we have observed a relation between the captured image intensity and the distance from the camera. Therefore, as a contrast, we attempt to exploit these scattering effects to obtain 3D depth cues. In order to learn the relation between the scattering effects and the depth, we utilize deep networks to help extract and build high-level features. In this paper, we propose a novel classification approach for depth map estimation of hazy images using deep learning.

Original languageEnglish
Title of host publicationProceedings - 4th Asian Conference on Pattern Recognition, ACPR 2017
PublisherInstitute of Electrical and Electronics Engineers Inc.
Pages244-249
Number of pages6
ISBN (Electronic)9781538633540
DOIs
Publication statusPublished - 13 Dec 2018
Event4th Asian Conference on Pattern Recognition, ACPR 2017 - Nanjing, China
Duration: 26 Nov 201729 Nov 2017

Publication series

NameProceedings - 4th Asian Conference on Pattern Recognition, ACPR 2017

Conference

Conference4th Asian Conference on Pattern Recognition, ACPR 2017
Country/TerritoryChina
CityNanjing
Period26/11/1729/11/17

Keywords

  • deep learning
  • depth
  • scattering media

Fingerprint

Dive into the research topics of 'Depth estimation for hazy images using deep learning'. Together they form a unique fingerprint.

Cite this