Development of AI Based Techniques for
Multispectral Satellite Image Dehazing
40130214 - SHARMILA. V
40130240 - UDHAYASURYA. G
PROJECT GUIDE
Dr. N. M. Nandhitha, M.E. , Ph.D.,
Professor and Dean,
School of EE
INTRODUCTION
• Satellite images provide wide information about the land/sea/oceans
• In recent years, satellite images are used extensively by researchers for every kind of
environmental prediction
• However the major challenge lies in the resolution of these images
• Hence considerable research is carried out to improve the quality of the satellite images
• One such preprocessing is dehazing whereby the satellite images are affected by smoke or fog
or smog
• In order to extract information from these images, it is necessary to develop dehazing
techniques for these satellite images
The proposed work aims at developing AI based image dehazing technique for
multispectral satellite images
LITERATURE SURVEY 1
Title: Single Image Dehazing using Multilayer Perceptron
Authors: Sebastian Salazar-Colores , Ivan Cruz-Aceves and Juan-Manuel Ramos-Arreguin
Problem Addressed: Outdoor images are susceptible to different atmospheric conditions, especially haze, fog, and smog.
The main effects caused by these atmospheric conditions are images with low contrast, distorted colors, and reduced visibility.
Consequently , the performance of systems, such as surveillance, target tracking, could be highly affected.
Methodology Proposed: The proposed method uses a multilayer perceptron to compute transmission map directly from the
minimum channel and a contrast stretching technique to improve the dynamic range of the restored image. The multilayer
perceptron is trained in terms of mean squared error using a training set of 80 images. To evaluate the restoration quality
compared with nine state-of-the-art dehazing methods. In addition, based on the average computational time achieved by the
proposed method, it can be highly suitable for real-time applications.
Conclusion: In this paper, a single image dehazing method was proposed. This method uses an artificial neuronal network
MLP to estimate the transmission map of a haze image. To obtain the optimal MLP configuration, a training set of 80 real-
world images was used. In this training experiments, a number of hidden layers containing different number of neurons were
tested, where the best performance in terms of the MSE was achieved using a 256- to 1024- MLP configuration with an MSE.
In order to evaluate the restoration quality of the proposed method achieves a superior performance than obtaining a PSNR
value and SSIM index over a test set of synthetic images. Given the suitable results in terms of restoration quality and
execution time of the proposed method with respect to the seven state-of-art dehazing methods, it can be highly appropriate to
be used in real-time systems.
3
LITERATURE SURVEY 2
Title: Single Image Dehazing Using Ranking Convolutional Neural Network
Authors: Yafei Song, Jia Li, Xiaogang Wang and Xiaowu Chen
Problem Addressed: In real-world scenarios, small particles suspending in the atmosphere (e.g., droplets and dusts) often
scatter the light. As a consequence, the clarity of an image would be seriously degraded, which may decrease the
performance of many multi media processing systems. Image enhancement methods can only alleviate this problem slightly. It
is still helpful to develop effective dehazing methods to recover the clear image from an input hazy or foggy image.
Methodology Proposed: This paper presents a novel Ranking Convolutional Neural Network (Ranking-CNN). In Ranking
CNN, a novel ranking layer is proposed to extend the structure of CNN so that the statistical and structural attributes of hazy
images can be simultaneously captured. By training Ranking CNN in a well-designed manner, powerful haze-relevant features
can be automatically learned from massive hazy image patches. Based on these features, haze can be effectively removed by
using a haze density prediction model trained through the random forest regression.
Conclusion: This paper presents a method to dehaze an image based on the features which are automatically learned from
massive 12 hazy images. The proposed image dehazing method, which is based on the features, also achieves satisfactory
results on synthetic and real world data. At the same time, as we extract features for every pixel, the weakness of our method
is its efficiency, which should be further improved in the future work, i.e., via adopt ing FCN framework [24] to reduce
redundant computations.
4
LITERATURE SURVEY 3
Title: Haze Removal of Multispectral Remote Sensing Imagery Using Atmospheric Scattering Model-Based Haze Thickness
Map
Authors: Hyeongseok Choi, Heunseung Lim, Soohwan Yu and Joonki Paik
Problem Addressed: The very high-resolution remote sensing images have been used for agriculture, geology, military, and
earth meteorology. However, the particles in atmosphere such as smog, dust, and vapor make the light scattering. That
saying, haze is affect at image low-contrast, signal saturation, losing color. To solve this problem, the haze removal methods
for the remote sensing imagery have been proposed in the literature.
Methodology Proposed: This paper presents a haze removal method for the multispectral remote sensing imagery. The
proposed method estimates the haze thickness map, which represents the amount of the haze. The haze removal is
performed using the atmospheric scattering model. Experimental results demonstrate that the proposed method can perform
haze removal without over dehazing and color distortion compared with existing methods.
Conclusion: This paper presented the haze removal method for multispectral remote sensing imagery. The proposed method
performs the haze removal using the atmosphere scattering model. Since the proposed method estimates the amount of haze
based on the wavelength, the resulting image shows better haze removal performance compared with existing methods. In
addition to the remote sensing imagery, the proposed method can be implemented to the digital image and videos acquired by
consumer unmanned aerial vehicle (UAV) and drones.
5
LITERATURE SURVEY 4
Title: Dehazing for Multispectral Remote Sensing Images Based on a Convolutional Neural Network With the Residual
Architecture
Authors: Manjun Qin, Fengying Xie , Wei Li, Zhenwei Shi and Haopeng Zhang
Problem Addressed: Multispectral remote sensing imagery can provide abundant ground information, which becomes an
essential way to explore earth resources and ecological environment. However, multispectral images are often affected by
atmospheric conditions such as haze, fog, and cloud. These phenomena reduce the image visibility and lead to the loss of
texture details, which bring obstacles for many applications such as land classification and target detection.
Methodology Proposed: A novel dehazing method based on a deep convolutional neural network (CNN) with the residual
structure is proposed for multispectral remote sensing images. First, multiple CNN individuals with the residual structure are
connected in parallel and each in dividual is used to learn a regression from the hazy image to the clear image. Then, the
outputs of CNN individuals are fused with weight maps to produce the final dehazing result. The designed network is end-to-
end, and putting a hazy image into it, the clear scene can be restored. To train the network, a wavelength-dependent haze
simulation method is proposed to generate labeled data, which can synthesize hazy multispectral images highly close to real
condi tions. Experimental results show that the proposed method can accurately remove the haze in each band of
multispectral images under different scenes.
Conclusion: Haze often appears in multispectral remote sensing images, which decreases the image visibility and causes
wrong interpretation. In this paper, a novel haze removal method based on the deep CNN is proposed for multispectral remote
sensing images. Thus, the learning difficulty is decreased. These individuals are trained using different levels of haze samples,
and they have different dehazing abilities. 6
LITERATURE SURVEY 5
Title: Single Image Dehazing via NIN-DehazeNet
Authors: Kangle Yuan and Naixue Xiong , Jianguo Wei , Wenhuan Lu
Problem Addressed: Image dehazing is a challenging problem in the eld of computer vision. The purpose of Image dehazing
is to recover a clear image from one single noisy frame caused by haze, fog or smoke. The dehazing algorithms have thus
been widely considered, as a challenging instance of (ill-posed) image restoration and enhancement. Similar to other
problems like image denoising and super-resolution earlier dehazing work assumed the availability of multiple images from the
same scene
Methodology Proposed: Singleimage dehazing has always been a challenging problem in the eld of computer vision.
Traditional image defogging methods use manual features. With the development of arti cial intelligence, the defogging
method based on deep learning has developed rapidly. In this paper, we propose a novel image defogging approach called
NIN-DehazeNet for single image. This method estimates the transmission map by NIN-DehazeNet combining Network-in-
Network with MSCNN(Single Image Dehazing via Multi Scale Convolutional Neural Networks).
Conclusion: In this paper we propose NIN-DehazNet, which estimates the transmission map by combining Network-in-
Network with MSCNN(Single Image Dehazing via Multi-Scale Convolu tional Neural Networks). And we compare it with some
methods which could get good effect about dehazing. The experiments show that the method we propose perform best and it
has the best effect about dehazing than other dehaz ing algorithms . The PSNR increased as well as the SSIM improved. The
experiments also show that our method is fast.
7
OBJECTIVES
PROBLEM STATEMENT
Proposed work aims at developing AI based dehazing techniques for multispectral satellite
images. Performance is measured in terms of contrast ratio, Signal to Noise Ratio and similarity
Index
THE DETAILED OBJECTIVES ARE:
• To collect multispectral haze satellite images
• To determine air-light and dark channel from the given image
• To estimate the input transmission channel
• To develop AI based technique for estimating the output transmission channel
• To determine the performance metrics
8
RESEARCH DATABASE (HAZED IMAGES)
The dataset contains 1200 images with various levels of thickness with its
corresponding ground truth images. These images are captured using SAR
(Synthetic Aperture Radar).
(a) Thick Haze Image (b) Moderate Haze (c) Thin Haze Image
Image
9
EXISTING WORK
MULTILAYER PERCEPTRON
• The existing method uses a multilayer perceptron to compute the transmission map
directly from the minimum channel and a contrast stretching technique to improve the
dynamic range of the restored image.
• The multilayer perceptron is trained in terms of mean squared error using a training set of
80 images. To evaluate the restoration quality compared with nine state-of-the-art
dehazing methods. In addition, based on the average computational time achieved by the
proposed method, it can be highly suitable for real-time applications.
10
TRAINING PROCESS OF MLP
11
PROPOSED METHODOLOGY
The proposed algorithm combines several key components such as
• Dark Channel Prior
• Air-Light estimation
• Transmission map Estimation
• Guided Filter Application
• Radiance estimation
• A Convolutional Neural Network
It appears to be a comprehensive dehazing algorithm designed to improve the
quality of hazy images. 12
BLOCK DIAGRAM OF THE PROPOSED SYSTEM
Air Light
Haze Image Input
Estimation
Initial Transmission
using the DCP
Refining the Initial
Transmission using
CNN and Guided Filter
Radiance Recovery Dehazing Unit Dehazed Image
13
DARK CHANNEL PRIOR ESTIMATION
Start
• The Dark Channel Prior is used to
estimate the thickness of the haze or Add Padding to the image
fog in an image.
Initialize the dark channel matrix
• By identifying the darkest regions
within an image, the algorithm can Initialize the loop to find the darkest pixel
estimate the lower bound of the haze,
which is proportional to the depth of Is the pixel No
darkest?
the scene Yes
End
14
AIR-LIGHT ESTIMATION
Start
• The Dark Channel Prior is used to estimate the Calculate Dark channel and get
image dimensions
thickness of the haze or fog in an image.
Determine the number of pixels to
• By identifying the darkest regions within an search
image, the algorithm can estimate the lower Sort the dark channel values and
initialize accumulator
bound of the haze, which is proportional to the
All the pixels No
depth of the scene. were
searched?
• Deep neural networks, such as Convolutional Yes
Neural Networks (CNNs), have been employed Estimate Air-light
to leverage the Dark Channel Prior in an end-to-
End
end dehazing framework. 15
TRANSMISSION MAP ESTIMATION AND REFINEMENT
• Transmission Map estimation is done by training the
Start
CNNs on pairs of hazy and haze-free images to learn
the complex mapping between them, allowing for the Initialize the image dimensions
estimation of the transmission map directly from the
Create Air-Light Mask
hazy input.
• Guided filtering is a technique that can be applied to Calculate transmission
refine the transmission map.
End
• It uses the guidance of the original hazy image to
enhance details while preserving edges. 16
RADIANCE ESTIMATION
• Radiance, denoted as L, measures the amount of Start
radiant flux (light energy) per unit solid angle, per unit
Initializing the image
projected area, and per unit wavelength. dimensions
• It is a fundamental property of light in a scene and Determining atmospheric
depends on the direction, position, and wavelength of parameters
the light rays. Finding the maximum
transmission
• Radiance is a key factor in scene understanding, as it
provides information about the intensity and distribution Calculating the radiance
of light in a scene.
End
17
CONVOLUTIONAL NEURAL NETWORKS
• It has been proven that Convolutional Neural Networks are effective in image dehazing
tasks as they learn intricate relationships and patterns directly from large datasets of
hazy and clear images.
• CNN models can capture contextual information in images, allowing them to consider
the relationships between different regions.
• Pre-trained models on large datasets for related tasks can be fine-tuned for image
dehazing.
18
PROPOSED CNN NETWORK
• The proposed network consists of three convolutional layers and each layer applies a
convolution operation to the input tensor with a specified number of filters.
• The first and second layer has 32 filters and the third layer has 64 filters. The filter size
is 3x3 for each layer.
• A flattened layer is added after the convolutional layers which converts the 3D output
tensor from the convolutional layers into a 1D tensor, which can be fed into fully
connected layers.
• Rectified Linear Unit (ReLU) activation function is used after each convolutional layer
and the first fully connected layer.
19
PROPOSED CNN NETWORK (Contd.)
• ReLU introduces non-linearity to the model and helps in learning complex patterns in
the data.
• A stride which is set to 1 is used for the convolution operation and Max-pooling layers
with a stride of 2 are applied after each convolutional layer, which reduces the spatial
dimensions of the feature maps by a factor of 2.
• The final layer uses softmax activation to produce class probabilities.
• The network is trained using the Adam optimizer with a learning rate of 1e-4
20
CNN NETWORK ARCHITECTURE
Estimated
Transmission
Map using DCP
Fully Connected
Convolutional Max Pooling A Ranking Layer Convolutional Max Pooling Layers
Layer Layer Layer Layer
Refined
Transmission
2x2 Map using CNN
3x3 3x3 2x2 3 Layers with 64
Non- Non-
Kernels Kernels neurons
Overlapping Overlapping
Windows Windows
21
TRAINING AND EVALUATION OF CNN
EPOCHS
• The CNN model is trained using 10 epochs and a batch size of 32 is used to process
the data.
• The choice of epochs and batch size affects how well the model learns from the data.
• By using 10 epochs, the model was allowed to learn from the dataset multiple times,
and the batch size of 32 ensured efficient processing.
LOSS
• The training loss and the validation loss of the model is almost similar. The training
loss is 0.135 and the validation loss is 0.1153.
ACCURACY
• The maximum accuracy was achieved after the iteration of 5th epoch. The training
accuracy is 0.952 and the validation accuracy is 0.9482. 22
TRAINING AND EVALUATION OF CNN (Contd.)
The below graphs show the loss and accuracy of the training and validation data in each
epoch.
23
GUIDED FILTER APPLICATION
• In the proposed system, a guided filter approach has been implemented specifically
tailored for the task of image dehazing, with the primary goal of improving visibility and
clarity within hazy scenes.
• This method relies on two crucial inputs: the guidance image, typically representing
the original hazy scene, and the filter image which is the hazy rendition of the scene to
be dehazed.
• The radius parameter permits control over the size of the local neighbourhood
considered during the filtering process, offering a delicate balance between haze
removal and preservation of fine details.
24
PERFORMANCE EVALUATION
COMPARISON OF VALUE OPTIMISED RESULTS
• The parameters such as the patch size of the dark channel prior and omega are
adjusted and tested with the images. The greater the patch size, the greater the rate
of dehazing.
• The parameter omega tells the amount of haze to be removed such that it will not
affect the original scene behind the haze.
• The omega is carefully increased in such a way that it should not erase the actual
scene behind the haze, leaving traces of the colour channel.
• It is set to 1 after evaluation and it is the maximum possible value for better dehazing.
25
RESULTS
Comparison of results haze based on Image Thickness
Thin haze Image
(b) Clear Image (c) Proposed Output (d) Existing Output
(a) Thin Haze
Moderate haze Image
(b) Clear Image (c) Proposed Output (d) Existing Output
(a) Moderate Haze
Thick haze Image
(b) Clear Image (c) Proposed Output (d) Existing Output
(a) Thick Haze
26
Comaparison of Images with different parameter
27
REFERENCES
1. Damian Wierzbicki, Michal Kedzierski and Aleksandra Sekrecka. A Method for Dehazing Images Obtained
from LowAltitudes during High-Pressure Fronts. Remote Sens. 2020, 12, 25; doi:10.3390/rs12010025
2. Akash Ramjyothi, Santonu Goswami. Cloud And Fog Removal From Satellite Images Using Generative
Adversarial Networks (GANs). 2021. hal-03462652
3. Jinjiang Li , Guihui Li , and Hui Fan. Image Dehazing Using Residual-Based Deep CNN. IEEE Trans. Vol-6,
May 2018.
4. G. Hari Hara Sudhan, R. Ganesh Aravind and K. Palani Thanaraj. Multispectral Analysis of Satellite Images
using Heuristic Algorithm. International Conference on Communication and Signal Processing, April 3-5, 2018,
India.
5. Zhijie He, Cailan Gong, Yong Hu, Fuqiang Zheng and Lan Li. Multi-Input Attention Network for Dehazing of Remote
Sensing Images. Appl. Sci. 2022, 12, 10523.
28
THANK YOU
29
