Description du projet
Une méthode optimisée pour éliminer la brume des images
En cas de mauvaises conditions météorologiques, comme en présence de brume, la qualité des images se dégrade fortement en raison de la présence dans l’air de particules en suspension qui diffusent la lumière. Les objets sont par conséquent difficiles à identifier, que ce soit par la vision humaine ou par les systèmes de vision par ordinateur. Les techniques actuelles d’élimination des effets de diffusion atmosphérique (débrumage) sont peu performantes pour les scènes de brume dense et homogène. Le projet NH-DEHAZE, financé par l’UE, développera des techniques de débrumage qui ne supposent pas une distribution homogène de la lumière et de la brume. Il établira le premier ensemble de données d’images comprenant des paires de scènes réelles, tant brumeuses que nettes. Des réseaux neuronaux profonds seront ensuite formés pour dériver des images claires à partir des images floues.
Objectif
In presence of haze, small floating particles absorb and scatter the light from its propagation direction. This results in selective and significant attenuation of the light spectrum, and causes hazy scenes to be subject to a loss of contrast and sharpness for distant objects. Besides, most computer vision and image processing algorithms (e.g. from feature extraction to objects/scene detection and recognition) usually assume that the input image is the scene radiance (haze-free image), and therefore strongly suffer from the color-shift, and low-contrast induced by hazy conditions. For instance, in normal visibility conditions the Traffic Sign Detection and Recognition (TSDR) module of the existing ADAS systems reaches a detection rate averaging around 90%, but drops below 40% in case of haze or poor illumination conditions1. Therefore, many recent works have explored inverse problem formulations and have designed dedicated image enhancement methods to address the dehazing problem. However, to estimate their key internal parameters (e.g. airlight in Koschmieder’s light transmission model), most of those solutions assume homogeneous distribution of light and haze, which is rarely the case in practice (e.g. lighting is non-uniform in space and frequency during the night, attenuation caused by haze depends on the light frequency).
Image dehazing thus remains a largely unsolved problem in case of dense and non-homogeneous haze scenes.
As a federating objective, our project aims at implementing dehazing methods that are suited to dense and non-homogeneous hazy scenes. This implies the following tasks:
(O1) build up the first (world-wide) image dataset including pairs of hazy and haze-free scenes, for which hazy scenes include real, dense, and non-homogeneous haze;
(O2) develop and train deep dehazing neural networks to derive the dehazed images from hazy inputs.
(O3) train deep image interpretation models that are suited to images captured in adverse conditions.
Champ scientifique
Programme(s)
Régime de financement
MSCA-IF - Marie Skłodowska-Curie Individual Fellowships (IF)Coordinateur
1348 Louvain La Neuve
Belgique