Efficient Image Dehazing Algorithms with Conventional and Deep Learning Based Techniques Using Multi-Spectral Images
Loading...
Date
2026
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Department of Electronics and Communication Engineering, NITS
Abstract
Image haze is a common problem in photography and imaging that occurs due
to particles like dust, smoke, and moisture in the atmosphere. These particles
scatter light, which leads to a reduction in visibility and a degradation in image
quality. This scattering effect reduces contrast and visual clarity. The impact
of haze is particularly significant in real-time applications, such as autonomous
driving, surveillance, and aerial photography, where the quality of visual data
is crucial for performance and accuracy. In autonomous driving, for instance,
clear images are essential for the vehicle’s vision systems to accurately detect and
interpret the surrounding environment. Haze can obscure important details like
road signs, pedestrians, and other vehicles, potentially leading to unsafe situations.
Similarly, in surveillance, clear imagery is critical for the accurate identification
of individuals and activities. Hazy images can hinder the effectiveness of these
systems, compromising security and safety.
To address these challenges, image dehazing techniques have been developed.
These techniques aim to restore the clarity and contrast of hazy images, making
them more useful for their intended applications. Image dehazing can be approached
using conventional methods as well as deep learning methodologies. Conventional
dehazing methods are often based on physical models of atmospheric scattering.
However, conventional methods have limitations, particularly in handling complex
and diverse haze conditions.
This thesis mainly focus on the development of dehazing algorithms, both
conventional and deep learning-based, has significantly advanced the field of
image processing. These algorithms not only improve the visual quality of images
but also enhance the performance of vision systems in various applications. By
addressing issues related to low contrast and haze, these techniques contribute to
safer autonomous driving, more effective surveillance, and higher-quality aerial
photography.
The first approach proposed in this thesis is a novel methodology for dehazing
single images using multi-scale image fusion. The process involves extracting
different scale images from a single hazy image, specifically a white-balanced
image and a luminance-based image. These images are then fused using
their corresponding weight maps. To avoid unwanted enhancements, pyramid
decomposition is applied to the weight maps and input images, enhancing contrast
and sharpening the image. This approach effectively removes haze, improving both
color and texture information, and produces clear images from a single hazy input.
A fast and efficient method is developed with an improved U-Net architecture
for image dehazing, enhancing images captured under low illumination, fog, and
snow conditions. The methodology involves replacing the pooling operations
in the traditional U-Net with Discrete Wavelet Transform (DWT) for up and
downsampling. Additionally, an attention module is introduced by fusing up and
down samples to identify and recover missing low-level features in the up-samples.
Another approach is introduced with a DeepNet convolutional network
architecture for effective haze removal from RGB images by accurately estimating
the transmission map, indicating light transmission at each pixel. The DeepNet
model leverages deep learning to predict the initial transmission map, capturing
essential features of haze-affected areas. This initial map is then refined using
Near-Infrared (NIR) images, which provide additional, non-visible information. A
multi-spectral transmission map fusion approach is employed to further enhance
accuracy, combining transmission maps from both RGB and NIR images. The
fusion process utilizes haze-aware weight generation, assigning appropriate weights
to different regions based on their haze levels, thus leveraging the strengths of each
spectral domain for a more detailed and accurate transmission map.
Finally a novel image dehazing framework is presented using a Generative
Adversarial Network (GAN) architecture, featuring a U-Net generator and dual
discriminators. The methodology involves leveraging the GAN’s capabilities to
transform hazy input images into clear, dehazed versions. The U-Net generator,
renowned for its efficiency in image-to-image translation tasks, plays a central role
in this process. Also, dual discriminators for two different spectral input (RGB
and NIR) images. It enhances the dehazing by effectively learning and removing
intricate haze patterns. The framework trains the GAN to generate high-quality,
perceptually realistic images, resulting in significantly improved image quality.
This advanced approach provides a robust and reliable solution for generating
clear, haze-free images from hazy inputs, crucial for outdoor vision applications.
The proposed image dehazing architectures were evaluated using various
publicly available datasets, including HazySky, I-Haze, O-Haze, See-in-the-Dark
(SID), Exclusively Dark Image Dataset (ExDARK), REalistic Single Image
DEhazing (RESIDE), and RGB-NIR. Additionally, the performance of low-light
image enhancement using an improved U-Net algorithm was analyzed with several
real-time images captured on the campus of the National Institute of Technology
Silchar. This study presents efficient and optimized approaches for image dehazing.