Image De-hazing for Vision Based Applications
Loading...
Files
Date
2022
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
NHCE
Abstract
Images that are captured in haze or fog condition may experience degradation, leading to a loss of accuracy in color, contrast, and visibility. This degradation occurs due to atmospheric particles that scatter and weaken the source radiation. The degree of degradation depends on various factors such as the density of environmental particles, their frequency, and their distance from the capturing device. The current techniques for dehazing images either rely on assumptions to reproduce the transmission map or use a learning system to estimate the dehazed image directly. A recent analysis of popular image dehazing methods using spectral dark images revealed that existing techniques perform poorly with frequency group selection and haze thickness levels. In this study, we propose an effective network called SPIDE-NET that uses spectral and prior-based image dehazing and enhancement techniques. Our approach outperforms existing methods by using spectral dark images from various frequency groups and haze thickness levels. The SPIDE-NET comprises two networks:
1) Spectral Image De-Hazing Network, which is trained on multi-spectral hazy images ranging from 450 nm to 720 nm, and
2) Multiscale Prior-based Image De-hazing Network that uses multi-scale dark channel and variance reduction priors on image trios selected from a multi-spectral hazy image dataset.