Project work on

IMPROVEMENT OF TEMPORAL RESOLUTION OF SCHLIEREN

MEASUREMENTS IN SUPERSONIC JET FLOW USING DATA-DRIVEN
METHODS

by

HARSHIT SOLANKI
(20117058)

ARPIT PRITHU
(20117031)

BHOOS ANUP SATISH

(20117039)

ADITYA SHANKAR SHUKLA

(20117012)

Under the supervision of

DR. NEETU TIWARI

PROF. B.K. GANDHI

MECHANICAL AND INDUSTRIAL ENGINEERING DEPARTMENT

INDIAN INSTITUTE OF TECHNOLOGY ROORKEE
ROORKEE – 247667 (INDIA)

MAY 2024
 DECLARATION
We hereby declare that the work carried out in this report entitled, “Improvement of
Temporal Resolution of Schlieren Measurements in Supersonic Jet Flow using
Data-Driven Methods” is presented for the course MIN-400 A/B (B.Tech. Project)
submitted to the Department of Mechanical and Industrial Engineering, Indian
Institute of Technology Roorkee (India), is an authentic record of our work carried
out under the supervision of Dr. Neetu Tiwari and Prof. B.K. Gandhi, MIED, IIT
Roorkee.

We have not submitted the record embodied in this project report for the award of any
other degree or diploma in any institute.

May 2024
IIT Roorkee

Harshit Solanki Arpit Prithu Bhoos Anup Satish Aditya Shankar Shukla
20117058 20117031 20117039 20117012

1
 CERTIFICATE

This is to certify that the report submitted by Mr Harshit Solanki, Mr Arpit Prithu, Mr
Bhoos Anup Satish and Mr Aditya Shankar Shukla on “Improvement of Temporal
Resolution of Schlieren Measurements in Supersonic Jet Flow using Data-Driven
Methods” is an authentic record of their project work which they have satisfactorily
completed under my supervision.

This report is submitted as a partial requirement for completing the MIN-400 B.Tech.
Project, which is necessary for the conferment of the bachelor's degree.

May 2024

Dr. Neetu Tiwari

MIED, IIT Roorkee

2
 ACKNOWLEDGEMENTS

This project would not have been possible without the kind support and help of our
supervisor, Dr. Neetu Tiwari. She supported us with her time, knowledge and
encouragement throughout the course of this project.

We also want to express our sincere gratitude to the thermal group committee
members Dr. Kirti Bhushan Mishra, Dr. Ajit Kumar Dubey and Dr. Alankrita Singh
for their edifying insights and feedback during the evaluations.

May 2024
IIT Roorkee

Harshit Solanki Arpit Prithu Bhoos Anup Satish Aditya Shankar Shukla
20117058 20117031 20117039 20117012

3
 TABLE OF CONTENTS

Chapter Title Page

Declaration 1
Certificate 2
Acknowledgements 3
List of Figures 7
List of Tables 9
Glossary 10
List of Symbols 11
1 Introduction 13
1.1 Motivation 13
1.2 Research context 14
1.3 Objectives 14
2 Extended Proper Orthogonal Decomposition 15
2.1 Proper orthogonal decomposition 15
2.2 Estimation of time-resolved velocity fields with extended 17
POD
2.2.1 Data extraction 17
2.2.2 Data pre-processing 17
2.2.3 Truncation and data compression 18
2.2.4 Finding common indexes and using extended modes 19
2.2.5 Reconstructing the snapshots 20
2.3 Results 22
3 Shallow Neural Networks 25
3.1 Architecture of Shallow Neural Networks 25

4
 3.1.1 Layered structure with interconnected artificial 25
neurons
3.1.2 Linear transformation followed by the non-linear 25
activation function
3.1.3 Forward pass 26
3.1.4 Backward pass or backpropagation 26
3.1.5 Learning rate 27
3.1.6 Regularisation 27
3.1.7 Optimisation 27
3.2 Linear autoencoders 28
3.2.1 Encoding 28
3.2.2 Decoding 28
3.2.3 Reconstruction 28
4 Convolutional Neural Networks 29
4.1 Architecture of convolutional neural networks 29
4.1.1 Convolutional layers 29
4.1.2 Filters 29
4.1.3 Feature maps 29
4.1.4 Pooling layers 29
4.1.5 Activation layers 30
4.1.6 Fully connected layers 30
4.1.7 Network architecture flow 30
4.2 U-net convolutional neural network 31
4.2.1 Contracting path (encoder) 31
4.2.2 Expansive path (decoder) 31
4.2.3 Skip connections 32
4.2.4 Benefits of U-Net 32
5
 4.2.5 Implementation of U-Net 33
5 Alternative Models for Better Reconstruction 34
5.1 General Adversarial Network 34
5.1.1 Architecture of GAN 34
5.2 Variational autoencoder 35
5.2.1 Working of VAE 35
5.2.2 Architecture of VAE 36
6 Results and Discussion 37
6.1 Reconstructed flow snapshots 37
6.2 Root mean square error (RMSE) 41
7 Conclusions 42
7.1 External proper orthogonal decomposition 42
7.2 Commentary on other models 42
8 Future Prospects 44
8.1 Combining CNN and LSTM networks 43
8.1.1 Capturing spatial features with CNN 43
4.1.2 Handling temporal dependencies with LSTM 44
4.1.3 Benefits of this approach 44
4.1.4 Challenges 44
9 References 45

6
 LIST OF FIGURES

Figure Description Page

Fig. 1 Proper orthogonal decomposition through economy-sized 17
matrix single-value decomposition
Fig. 2(a) Non-time-resolved fluid flow snapshot 18
Fig. 2(b) Time-resolved fluid flow snapshot 18
Fig. 3(a) Original non-time-resolved fluid flow snapshot 18
Fig. 3(b) Fluid flow snapshot after contraction along the horizontal 18
direction
Fig. 4 Fluid flow snapshot following mean subtraction 19
Fig. 5(a) Time-resolved truncated snapshot 20
Fig. 5(b) Non-time-resolved truncated snapshot 20
Fig. 6(a) Non-time-resolved truncated snapshot (reproduced) 21
Fig. 6(b) Reconstructed snapshot for deltr = 23 21
Fig. 7 Reconstruction difference 21
Fig. 8 Reconstructed snapshot for deltr = 10 22
Fig. 9 Reconstruction error vs number of virtual modes considered 22
Fig. 10 Percentage energy represented vs number of modes 23
considered (for time-resolved snapshots)
Fig. 11 Percentage cumulative energy vs number of modes considered 23
Fig. 12 Percentage reconstruction error vs Nth non-time-resolved 24
snapshots
Fig. 13 Architecture of ANN 25
Fig. 14 Schematic of linear autoencoder 29
Fig. 15 Architecture schematic of the convolutional neural network 32
used in this study
Fig. 16 Architecture schematic of the U-Net used in this study 33

7
Fig. 17 Schematic of GAN architecture 35
Fig. 18 Original fluid flow snapshot 38
Fig. 19 Reconstructed fluid flow snapshot using EPOD 38
Fig. 20 Reconstructed fluid flow snapshot using ANN 39
Fig. 21 Reconstructed fluid flow snapshot using CNN 39
Fig. 22 Reconstructed fluid flow snapshot using linear autoencoder 40
Fig. 23 Reconstructed fluid flow snapshot using U-Net CNN 40
Fig. 24 Reconstructed fluid flow snapshot using CNN-VAE 41
Fig. 25 Reconstructed fluid flow snapshot using GAN 41
Fig. 26 Schematic representation of CNN-LSTM hybrid network 43

8
 LIST OF TABLES

Table Description Page

Table 1 RMSE values of different techniques/methodologies 41

9
 GLOSSARY

Abbreviations Meaning
EPOD Extended Proper Orthogonal Decomposition
CNN Convolutional Neural Network
ANN Artificial Neural Network
ReLU Rectified Linear Unit
GAN Generative Adversarial Network
VAE Variational Autoencoder

10
 LIST OF SYMBOLS

Symbol Description
U Snapshot matrix
𝛹 Orthogonal matrix of dimensions (𝑚, 𝑚), where 𝑚 represents the
number of spatial resolution points of a particular snapshot
𝛴 Singular diagonal matrix of dimensions (𝑚, 𝑛) wherein the
non-zero elements along its diagonal explain the proportion of
energy inherent in the original snapshots
𝛷 Orthogonal matrix of dimensions (𝑛, 𝑛) where n represents the
number of snapshots
X Snapshot matrix after truncation
𝛯 Temporal correlations matrix
LN Nth layer of a neural network
J The cost function of a neural network
W Weights associated with neurons in the layers of a neural network
α Hyperparameter that controls the learning rate
λ Hyperparameter used for weight decay, also called regularisation

Superscript Description
’ Reconstructed matrix
T Transpose of matrix
i ith entity

Subscript Description
tr Time-resolved matrix
ntr Non-time-resolved matrix
pr Predicted time-resolved matrix
11
est Estimated matrix
pred Value predicted by the neural network
true True value

12
 Chapter 1
INTRODUCTION

Accurate fluid flow simulation is crucial in various engineering fields, including fluid
dynamics, aerospace, automotive, and civil engineering. Traditional experimental
methods for flow simulation are often costly, time-consuming, and limited in
capturing complex flow phenomena accurately. In recent years, data-driven
techniques have emerged promising to improve accurate flow simulation by
leveraging computational power and large datasets.

1.1 Motivation
Capturing flow dynamics is a complex task that requires precise and rapid data
acquisition. A critical factor in this process is the sampling rate, which refers to the
frequency at which data points are recorded over time. A high sampling rate is crucial
for flows involving extremely high-speed phenomena to accurately capture the flow
field's dynamic behaviour.

A sampling rate of 100,000 samples per second (105 Hz) is typically necessary to
adequately capture the rapid changes and intricate details in supersonic flow
phenomena. This high sampling rate ensures the camera can record enough data
points within a given time frame to represent the flow dynamics accurately.

However, despite their advanced capabilities, modern high-speed cameras often need
to be improved regarding sampling rates. For example, a common resolution for
high-speed cameras might be 1024 x 1024 pixels, capturing images composed of
1,048,576 individual pixels. Despite this high resolution, the sampling rate of such
cameras typically ranges around 10,000 samples per second (10 kHz), significantly
lower than the ideal rate for capturing flow phenomena.

This discrepancy between the required sampling rate for flow capture and the
practical limitations of modern high-speed cameras arises from various design
constraints inherent to these devices. One crucial constraint is the trade-off between
sampling rate and field of view. Increasing the sampling rate typically requires
sacrificing the captured images' field of view or spatial resolution. Conversely,
maximising the field of view often comes at the expense of the sampling rate.

This trade-off exists due to limitations in camera sensor technology, data processing
capabilities, and physical constraints related to capturing and storing high-speed data.
Another challenge associated with using sensors is their intrusive nature. This means

13
that their installation can perturb the flow properties around the point of installation,
potentially leading to inaccuracies in property measurements. Additionally, sensors
are expensive and cannot be deployed comprehensively throughout the flow to study
their properties.

1.2 Research context

In recent years, high-speed Photron cameras have become increasingly prevalent as
an alternative to traditional sensors for analysing fluid flows. This approach involves
strategically placing multiple cameras at various points of interest within the flow
field. The cameras continuously capture snapshots, which are then utilised to
reconstruct and study fluid flow patterns. The camera-based method offers several
advantages, including cost-effectiveness, user-friendliness, and a non-intrusive
nature, as the cameras do not interfere with the flow. These benefits have collectively
contributed to the growing popularity of high-speed Photron cameras, making them a
preferred alternative to traditional sensor technologies for fluid flow analysis.

1.3 Objectives
The prerogative is to develop a method that increases the temporal resolution to the
desired value while keeping a high spatial resolution without relying on extensive
deployment of sensors. The project strives to advance fluid flow simulation
techniques by harnessing the power of various data-driven techniques. The primary
objectives of this project work are listed below:

1. Reconstruction of high-resolution supersonic fluid flow from

low-resolution snapshots: The primary aim is to develop a method to
reconstruct high-resolution images of supersonic fluid flow from
low-resolution images. This process aims to enhance the quality and detail of
the flow snapshots, allowing for better visualisation and analysis of the
intricate details present in supersonic flow phenomena. To achieve this, the
project will employ data-driven and machine-learning techniques, which can
leverage the inherent patterns and relationships within the low-resolution data
to reconstruct high-resolution flow representations.

2. Analyse and compare reconstruction methods: The project aims to analyse

and compare the performance of two different reconstruction methods: the
EPOD (Extended Proper Orthogonal Decomposition) approach and machine
learning-based methods. By comparing the reconstructed flows obtained from
these two approaches, the project seeks to determine which method yields
better accuracy, computational efficiency, or other metrics.
14
 Chapter 2
EXTENDED PROPER ORTHOGONAL DECOMPOSITION

This research presents an innovative approach for estimating fluid velocities over
time} termed Extended Proper Orthogonal Decomposition (EPOD). EPOD leverages
data from synchronised cameras observing fluid flow dynamics to analyse temporal
changes, providing insights into fluid behaviour beyond spatial patterns.
Demonstrated efficacy in diverse scenarios, including jet flows and flows with
obstacles, underscores EPOD's robustness as a foundational tool in fluid dynamics
analysis.

In this study, we introduce a novel methodology that amalgamates two distinct types
of measurements to enhance understanding of fluid behaviour: (a) low-resolution,
time-resolved images capturing fluid flow dynamics, and (b) high-resolution,
non-time-resolved camera data providing detailed spatial information. By integrating
these measurements, we aim to improve the accuracy of fluid density gradient
estimation. Notably, our approach incorporates a mechanism to address time delays
inherent in the data fusion process. This enables a more comprehensive
characterisation of temporal evolution in fluid density, facilitating more profound
insights into fluid dynamics.

2.1 Proper Orthogonal Decomposition

This section explains the data organisation and implementation intricacies of a

method rooted in Extended Proper Orthogonal Decomposition (EPOD), tailored
explicitly for estimating pressure gradients from snapshot data.

We start by considering a set of 𝑛 snapshots representing the dynamic fluctuations

within the pressure gradient field. These snapshots are systematically arranged into a
matrix denoted as 𝑈, wherein each column corresponds to a distinct snapshot.

To effectively capture the essential patterns within the dataset and facilitate
dimensionality reduction, we leverage an economy-size Singular Value
Decomposition (SVD) on matrix 𝑈. This SVD operation dissects 𝑈 into three core
constituent matrices: 𝛹, 𝛴, and 𝛷.

(2.1)

15
𝛹 represents an orthogonal matrix of dimensions (𝑚, 𝑚), where 𝑚 characterises the
spatial resolution points encapsulated within each snapshot. 𝛴 manifests as a singular
diagonal matrix of dimensions (𝑚, 𝑛), wherein the non-zero elements along its
diagonal delineate the proportion of inherent energy within the original snapshots. 𝛷,
conversely, manifests as another orthogonal matrix of dimensions (𝑛, 𝑛), with 𝑛
signifying the total number of snapshots.

We discern and retain the most pertinent modes of interest to streamline the data's
dimensionality. This involves selecting the top 𝑘 modes by extracting the first 𝑘
columns from matrices 𝛹 and 𝛷. Such a selective approach enables the capture of
predominant variability modes within the pressure gradient field, thereby reducing
the overall data size.

(2.2)

This process yields a reconstructed matrix, denoted as 𝑈’, with dimensions (𝑚, 𝑛),
while matrices 𝛹’, 𝛴’, and (𝛷T)’ possess dimensions (𝑚, 𝑘), (𝑘, 𝑘), and (𝑘, 𝑛)
respectively. This systematic reduction in dimensionality is instrumental as it
furnishes a concise yet informative representation of the original snapshot data. Such
a methodological framework finds applicability across diverse domains, including
fluid dynamics, structural mechanics, and signal processing, wherein extracting
salient features from high-dimensional datasets is paramount.

This process is valuable because it reduces the dimensionality of the original snapshot
data, providing a concise yet informative representation. Such an analytical
framework can be applied across various domains, including fluid dynamics,
structural mechanics, and signal processing, where extracting relevant features from
high-dimensional datasets is crucial.

16
 Fig. 1 Proper orthogonal decomposition through an economy-sized
matrix single-value decomposition (David W. Ashmore et al. (2022))

2.2 Estimation of time-resolved velocity fields with extended POD

2.2.1 Data extraction

The experimental setup consists of two high-speed Photron cameras that capture
snapshots of supersonic jet flows within a defined time frame, with data stored in
MRAW format. From these files, matrix data corresponding to both high-resolution
and low-resolution snapshots was extracted. Notably, the frequencies of
time-resolved and non-time-resolved snapshots inherently differed. The matrix
captured snapshots in both scenarios' x and y directions, while the z direction
portrayed temporal variations. Subsequently, this three-dimensional matrix was
reshaped into a two-dimensional matrix, where each column represented a snapshot at
a specific time.

17
 (a) (b)

Fig. 2(a) Non-time-resolved fluid flow snapshot and (b) Time-resolved fluid flow
snapshot

2.2.2 Data pre-processing

The snapshots were visualised by plotting images derived from their corresponding
matrices. Reconstruction efforts focused on capturing the maximum density gradient
variation. Time averaging was computed at each spatial position and subtracted for
time-resolved and non-time-resolved flow data.

The figure below presents time-resolved snapshots originating from three distinct
spatial locations.

(a) (b)

Fig. 3(a) Original non-time-resolved fluid flow snapshot and (b) Fluid flow snapshot
after contraction along the horizontal direction

18
 Fig. 4 Fluid flow snapshot following mean subtraction

2.2.3 Truncation and data compression

After obtaining both matrices from the above steps, the Singular Value
Decomposition (SVD) is carried out using the econ algorithm for faster and better
results.

(2.3)

(2.4)

The subscripts tr and ntr denote time-resolved and non-time-resolved flow fields,
respectively. Using the proper orthogonal decomposition (POD), we represent the
matrices using ‘k’ dominant modes.

(2.5)

(2.6)

Matrices after truncation can be represented as,

(2.7)

(2.8)

19
 Fig. 5(a) Time-resolved truncated snapshot and (b) Non-time-resolved truncated
snapshot

2.2.4 Finding common indexes and using extended modes

For a given time interval, the number of snapshots captured by the time-resolved and
non-time-resolved cameras are proportional to their frequencies. So, we prepare the
time instances at which both time and non-time-resolved snapshots are captured.

Now, we finalise the extended modes (extra snapshots) taken with each common
snapshot for reconstruction.

Maximum number of extended modes = (nt/nn) – 1 (2.9)

Here, nt and nn are the number of time-resolved and non-time-resolved snapshots,

respectively.

For each index i, we consider snapshots of X’tr from index i to (i + deltr), where deltr
is the number of extra snapshots considered. We combine these snapshots into a
single-column vector and store it in the ith column of another matrix named ycol.

20
For each index i considered, we consider snapshots of X’tr from index i to (i + deltr)
and combine them into a single-column vector and store them in the ith column of
another matrix named PP, where PP is a zero matrix of size (ps﹡(deltr + 1), nt –
deltr).

2.2.5 Reconstructing the snapshots

The extended POD modes 𝛷𝑒 corresponding to the non-time-resolved domain can be
estimated by projecting the non-time-resolved snapshot matrix 𝑈 onto the temporal
modes of the time-resolved snapshots.

(2.10)

The last two terms contain information about the temporal correlation between
non-time-resolved and time-resolved modes.

Knowing the POD spatial modes and singular values of the density gradient and of
the time-resolved snapshot matrix, as well as the temporal correlations matrix 𝛯, it is
possible to estimate the density gradient at a generic instant from a time-resolved data
snapshot PP sampled at that instant:

(2.11)

(a) (b)

Fig. 6(a) Non-time-resolved truncated snapshot (reproduced) and

(b) Reconstructed snapshot for deltr = 23

21
 Fig. 7 Reconstruction difference

Fig. 8 Reconstructed snapshot for deltr = 10

If observed carefully, the reconstructed image for deltr = 23 is a better resemblance of

the non-time-resolved snapshot with respect to the image where deltr = 10.

22
2.3 Results

Fig. 9 Reconstruction error vs number of virtual modes considered

Fig. 10 Percentage energy represented vs number of modes considered (for

time-resolved snapshots)

23
 Fig. 11 Percentage cumulative energy vs number of modes considered

Fig. 12 Percentage reconstruction error vs Nth non-time-resolved snapshots

24
 Chapter 3
SHALLOW NEURAL NETWORKS

Shallow neural networks are a widely used deep learning technique for fluid flow
reconstruction. These networks are attractive due to their ease of use, fast training
times, and ability to avoid overfitting the data. Unlike deeper networks, shallow
neural networks typically have three layers: an input layer, a single hidden layer, and
an output layer. This more straightforward structure offers several benefits. Firstly, it
keeps the number of parameters in the network low. This translates to less
computational power and memory needed for training, making them ideal for
working with limited datasets. The following section will explore how a shallow
neural network can be a viable alternative to existing reconstruction methods.

3.1 Architecture of Shallow Neural Network

3.1.1 Layered structure with interconnected artificial neurons

Fig. 13 Architecture of ANN

Imagine a stack of mathematical functions. Each layer represents a level of

abstraction. The first layer, the input layer, receives raw data. In our case, it is the
fluid density at each point. Subsequent layers (hidden layers) process and transform
this data. Neurons within each layer are interconnected, meaning the output from one

25
neuron becomes the input to others in the next layer. This chain reaction allows the
network to progressively extract features and higher-level patterns from the data.

Our shallow neural network consists of 4 neural layers: an input layer, an output
layer, and two hidden layers consisting of 100 and 400 neurons, respectively. The
input layer has 128 x 120 neurons, while the output layer consists of 1024 x 401
neurons.

Layers are denoted by LN, where L0 is the input layer, LH is the output layer and L1 to
LH–1 are the hidden layers.

(2.12)

3.1.2 Linear transformation followed by non-linear activation function

Within each neuron, a linear transformation is performed. This essentially means
multiplying each input by weight and summing them all up. Weights act as
coefficients, determining the influence of each input on the neuron's output. However,
a simple summation wouldn't allow the network to learn complex patterns. Here's
where the activation function comes in. It introduces non-linearity, acting like a filter
or threshold on the summed input. This allows the neuron to decide how strongly it
should "fire" its output based on the weighted sum.

We have selected ReLU as the activation function, as shown below.

(2.13)

We have selected ReLU (rectified linear unit) for the activation function for the
following reasons:

1. Computational efficiency: ReLU computes much faster compared to tanh or

sigmoid functions. This efficiency translates to faster training times and faster
predictions during inference.

2. Vanishing gradient problem: The vanishing gradient problem can occur in

deep neural networks. This happens when gradients become very small as they
propagate backwards through the network during training. Small gradients
make updating weights in earlier layers difficult, hindering the network's
learning ability. ReLU helps alleviate this problem because, for positive inputs,
the gradient remains constant (= 1), allowing for proper weight updates in
earlier layers.

26
3.1.3 Forward pass
This refers to how information travels through the network during training. The input
vector (array of data points) enters the network at the input layer. Each neuron in the
layer performs the weighted sum and applies its activation function. The resulting
outputs become the inputs for the next layer, and so on, until the final output layer
produces the network's prediction. It's like passing a baton in a relay race, with each
layer progressively transforming the information.

3.1.4 Backward pass or backpropagation

This is where the learning magic happens! The error signal is propagated backwards
through the network, layer by layer. Here's the key: the weights are adjusted
proportionally to their contribution to the overall error. Imagine slowly tweaking each
weight based on how much it affected the final output error. Backpropagation allows
the network to learn from its mistakes and gradually improve its ability to map inputs
to the desired outputs.

Cost function (J) is defined as:

(2.14)

(2.15)

3.1.5 Learning rate

This value controls how much the network should adapt to each training example. A
lower learning rate leads to more gradual weight adjustments, preventing the network
from overfitting (memorising the training data without generalising well to unseen
data). It's like setting the pace of learning – too fast, and the network might miss
important patterns; too slow, and it might take forever to learn.

3.1.6 Regularisation
In machine learning, regularisation encompasses techniques employed to combat
model overfitting. Overfitting occurs when a model excessively aligns with the
training data, memorising specific details and noise rather than capturing the
underlying generalisable patterns. This can lead to poor performance on unseen data.

27
Regularisation techniques achieve their goal by introducing a penalty term to the cost
function. The cost function quantifies the model's prediction error on the training
data. Regularisation modifies this function by adding a term penalising the model's
complexity. This complexity penalty can be based on the model's parameters (weights
and biases) or the overall structure.

We have used L2 normalisation for our experiment. L2 regularisation penalises the

model based on its weights' L2 norm (squared norm). The penalty term is λ ,
where λ is the hyperparameter and represents the sum of squares of all weights.

L2 regularisation shrinks all weights towards zero but doesn't necessarily eliminate
them. This reduces the overall magnitude of the weights, making the model less
sensitive to specific features and promoting smoother decision boundaries

3.1.7 Optimisation
They are algorithms that iteratively update the weights and biases within the network
to minimise a cost function (also known as a loss function). We have used Adam as
our optimiser with a learning rate of 0.001 and a decay rate of 0.9

3.2 Linear autoencoders

Linear autoencoders are a specific type of autoencoder architecture within neural
networks. An autoencoder is a neural network architecture designed to learn efficient
representations of input data.

3.2.1 Encoding
The input image is fed into the encoder part of the autoencoder. The encoder uses
linear functions (activations) to compress the image data into a lower-dimensional
representation called the latent space. In our study, we used ReLU as the activation
function; the input layer has 15,360 neurons, which decreases for the encoder part.

3.2.2 Decoding
The compressed representation from the latent space is passed to the decoder. The
decoder uses linear functions again to try and reconstruct the original image based on
the information in the latent space. Its output layer has 410,624 neurons reshaped to
the matrix of dimensions 1024*401 to reconstruct the original flow.

28
3.2.3 Reconstruction
The decoder aims to recreate the image by leveraging the learned linear relationships
between pixels in the original image.

Fig. 14 Schematic of linear autoencoder

29
 Chapter 4
CONVOLUTIONAL NEURAL NETWORK

Convolutional Neural Networks (CNNs) are a powerful type of neural network

architecture specifically designed for image, video, and time series data analysis.
Unlike standard neural networks that treat data as a flat vector, CNNs exploit the
inherent spatial structure of this data type.

4.1 Architecture of convolutional neural networks

4.1.1 Convolutional layers
Convolutional Neural Networks (CNNs) rely on a crucial component known as
convolutional layers. These layers execute a mathematical operation called
convolution to extract essential features from the input data. In this operation, a
compact filter, also known as a kernel, moves across the input data, calculating the
dot product between the filter elements and the corresponding data elements at each
position. Convolutional layers discern localised patterns within the data through this
iterative process, facilitating effective feature extraction.

4.1.2 Filters
Each filter learns to detect specific features, like edges, shapes, or colours. Multiple
filters are used within a convolutional layer to capture various features at different
scales and orientations.

4.1.3 Feature maps

The output of a convolutional layer is called a feature map. Each feature map
represents the activation of a specific filter across the entire input. The network learns
various feature maps, progressively building a hierarchy of increasingly complex
features.

4.1.4 Pooling layers

These layers downsample the feature maps produced by convolutional layers. This
reduces the dimensionality of the data, making it computationally cheaper and
potentially less prone to overfitting. Pooling layers typically use a pooling function
(like max pooling or average pooling) to summarise the information within a small
region of the feature map.

30
4.1.5 Activation layers
These layers introduce a crucial element of nonlinearity into the network. Activation
functions like ReLU (Rectified Linear Unit) and sigmoid are frequently employed for
this purpose. This nonlinearity is essential as it empowers the network to grasp
intricate connections between features, enabling it to make precise predictions.

4.1.6 Fully connected layers

In the final stages of the network, fully connected layers operate similarly to standard
neural networks. Here, neurons from all the previous layers are connected to every
neuron in the next layer. These layers take the high-level features extracted from the
convolutional and pooling stages and combine them to make the final prediction (e.g.,
image classification and object detection).

4.1.7 Network architecture flow

1. Input layer: The network receives the input data, which can be an image
represented by a 3D tensor (width, height, channels - RGB for colour images).
In this study, it is a time-resolved snapshot, a 3D tensor with dimensions 128 x
120 x 1 as it is in grayscale.

2. Convolutional layers: The input data goes through multiple convolutional

layers, each with its own set of filters. These layers extract features at different
levels of complexity. The kernel size is set to 3 x 3 for this experiment.

3. Pooling layers: Pooling layers are often inserted between convolutional layers
to reduce dimensionality and control overfitting.

4. Upsampling layers: Upsampling layers are a crucial component in

convolutional neural networks (CNNs) specifically designed to increase the
spatial resolution (width and height) of feature maps. In our study, we used
UpSampling2D layers, which take a feature map as input and increase its
spatial resolution through a specified upsampling factor. It achieves this by
duplicating existing pixel values (similar to nearest neighbour upsampling) or
using a basic form of bilinear interpolation.

5. Activation layers: After each convolution or pooling layer, an activation layer

introduces non-linearity, allowing the network to learn more intricate patterns.
We have used the ReLU activation function for our experiment.

6. Flattening: After the convolutional and pooling stages, the data is typically
flattened into a single vector before feeding into fully connected layers.
31
 7. Fully connected layers: These layers perform classification or regression tasks
based on the extracted features. The final output layer produces the network's
prediction. The fully connected output layer has 410,624 (1024 x 401) neurons,
which are later reshaped to a matrix of dimensions 1024 x 401 to get the
desired reconstructed flow.

Fig. 15 Architecture schematic of the convolutional neural network used in this study

4.2 U-Net convolutional neural network

U-Net, introduced in the paper "U-Net: Convolutional Networks for Biomedical
Image Segmentation" by Olaf Ronneberger et al., is a robust convolutional neural
network (CNN) architecture specifically designed for image segmentation tasks.
While initially developed for the medical field, its effectiveness has led to its
application in various segmentation problems beyond the field of medicine.

U-Net leverages an encoder-decoder structure with skip connections to achieve

accurate segmentation.

4.2.1 Contracting path (encoder)

This path focuses on extracting contextual information from the input image. It uses
standard convolutional layers with pooling operations to progressively reduce the
spatial resolution (width and height) while increasing the number of feature channels.
This allows the network to capture higher-level features from the data.

4.2.2 Expansive path (decoder)

This path aims to recover spatial resolution and produce a precise segmentation map.
It utilises upsampling techniques like transposed convolutions (deconvolutions) and

32
combines high-level features from the encoder with localised features from earlier
stages. This combination helps maintain precise object boundaries in the
segmentation output.

4.2.3 Skip connections

These are the secret sauce of U-Net. They directly connect corresponding layers in
the encoder and decoder paths. By copying feature maps from the encoder and
concatenating them with the upsampled feature maps in the decoder, skip connections
ensure that the decoder can access the precise spatial information from the earlier
encoder stages. This is crucial for accurate segmentation, especially when dealing
with intricate details and boundaries.

4.2.4 Benefits of U-Net

1. Effective with limited data: Skip connections and the encoder-decoder
approach enable U-Net to learn effectively even with limited training data
compared to standard CNN architectures.

2. Precise segmentation: U-Net excels at producing accurate segmentation

masks with well-defined object boundaries by leveraging high-level and
localised features through skip connections.

Fig. 16 Architecture schematic of the U-Net used in this study

33
4.2.5 Implementation of U-Net
In our study, we employed a U-Net network to refine further the reconstruction of
non-time-resolved data obtained using the EPOD method. This step aims to achieve a
lower reconstruction error.

First, the input (fluid flow reconstructed using EPOD) passes through multiple
convolutional layers that are responsible for dimensionality reduction and feature
extraction. As shown in the above figure, the input dimensions are reduced as they
pass through the encoder.

Then, the encoded data passes through the decoder part, which consists of the
deconvolution or upsampling layer. In this study, we have used a transposed
convolution layer, also known as a Conv2DTranspose. Unlike a standard
convolutional layer that extracts features from an input, a Conv2DTranspose layer
upsamples the input feature maps and learns filters to increase the data's spatial
resolution (width and height). This makes them particularly useful for tasks needing
to recover or even generate high-resolution outputs.

34
 Chapter 5
ALTERNATIVE MODELS FOR BETTER RECONSTRUCTION

5.1 Generative Adversarial Network

A General Adversarial Network (GAN) is a type of deep learning architecture that
involves two neural networks competing against each other in a zero-sum game
framework. The two networks are:

1. Generator network: This network generates new synthetic data instances

(e.g., images) by capturing the underlying distribution of the training data.

2. Discriminator network: This network evaluates the instances produced by the

generator and tries to distinguish between the real training data and the
synthetic data generated by the generator.

Fig. 17 Schematic of GAN architecture

The generator tries to generate indistinguishable data from real data, while the
discriminator tries to identify between real and generated data accurately. This
adversarial training process continues until the generator learns to produce synthetic
data that is highly realistic and can fool the discriminator.

5.1.1 Architecture of GAN

Generator model

The generator network is a deep neural network designed to take in a random noise
vector and produce a synthetic data sample, such as an image, audio, or text.
Typically, the architecture of the generator network comprises multiple layers of

35
upsampling or transposed convolutional layers, which are then followed by
non-linear activation functions.

The generator's objective in a GAN is to produce synthetic samples that are realistic
enough to fool the discriminator. The generator achieves this by minimising its loss
function JG​. The loss is minimised when the log probability is maximised.

(2.16)

Discriminator model

The discriminator in a Generative Adversarial Network (GAN) is a binary classifier

that evaluates both real data samples from the training dataset and synthetic (fake)
data samples generated by the generator. It assigns a probability score to each sample,
indicating how likely it is to be real or fake.

The adversarial training process aims to improve the discriminator's ability to

correctly identify the fake samples produced by the generator as fake and the real
samples from the training data as real.

During training, the generator and discriminator networks compete against each other.
As the discriminator becomes better at distinguishing between real and fake samples,
the generator must improve its ability to generate more realistic synthetic data to fool
the discriminator.

The discriminator's objective is to reduce the negative log-likelihood of correctly

distinguishing between generated (fake) samples as fake and real samples as real.
This loss function incentivises the discriminator to classify the generated samples as
fake precisely and the real ones as genuine, as expressed by the following equation:

(2.17)

5.2 Variational Autoencoder (VAE)

A Variational Autoencoder (VAE) is a neural network for learning and generating
new data, especially images. It's a form of generative model that can create new
images that look like they came from a dataset it was trained on.

The "variational" part of VAE comes from how it learns the latent space. Instead of
forcing the encoder to produce a single fixed point in the latent space for each input,
VAEs learn to create a probability distribution (like a range or spread) for each input.
This approach makes the model more flexible and robust.
36
5.2.1 Working of VAE
1. Encoder: The VAE starts with an encoder network that takes an input (like an
image) and converts it into a compact representation called a "latent space."
Think of this latent space as a compressed version of the input data.

2. Latent space: This compressed representation in the latent space captures the
essential features of the input data but in a more concise form. It's like
summarising a long story into a short paragraph.

3. Decoder: The decoder network then takes this compact representation from the
latent space and tries to reconstruct the original input data (like an image). It's
like expanding that short paragraph back into the whole story.

5.2.2 Architecture of VAE

Encoding side

1. Input layer: This layer takes the input data (e.g., an image) and passes it
through the subsequent layers.

2. Hidden layers: Process the input and produce the mean and variance of the
latent space distribution. The encoder produces these mean and variance,
defining the latent space's probability distribution.

3. Sampling layer: Takes the mean and variance from the encoder to sample a
point from the latent space distribution. This sampled point represents the
compact representation of the input data.

Decoding side

1. Input layer: This layer takes the sampled latent vector (z) as input.

2. Dense/upsampling layer: Depending on the input data type, the decoder may
use dense layers or upsampling layers (or a combination of both) to gradually
increase the latent representation's dimensionality.

3. Convolutional/dense layer: Similar to the encoder, the decoder may use

convolutional layers or dense layers (or a combination of both) to process the
upsampled features and generate the reconstructed output.

4. Output layer: The final layer of the decoder produces the reconstructed
output, which ideally should match the original input data as closely as
possible.

37
 Chapter 6
RESULTS AND DISCUSSION

6.1 Reconstructed flow snapshots

6.1.1 Original flow snapshot

Fig. 18 Original fluid flow snapshot Fig. 19 Reconstructed fluid flow

snapshot using EPOD

6.1.2 Using EPOD

EPOD can capture temporal relations, which isn’t possible through ANN since it only
considers spatial relations during reconstruction. The reconstructed flow snapshot
using EPOD is shown in Fig. 18 above.

6.1.3 Using ANN

As evident, ANN is not suitable for image reconstruction purposes. As image
resolution and complexity increase, the computational and memory requirements of
ANNs can grow rapidly, making it challenging to scale up for high-quality image
reconstruction.
38
 Fig. 20 Reconstructed fluid flow Fig. 21 Reconstructed fluid flow
snapshot using ANN snapshot using CNN

6.1.4 Using CNN

Although CNNs are primarily designed for image recognition and classification tasks,
which aim to extract relevant features from the input image, image reconstruction
requires explicitly encoding spatial information and details, which CNNs do not
explicitly learn or preserve. The reconstructed flow snapshot using CNN is shown in
Fig. 20 above.

6.1.5 Using linear autoencoder

Linear autoencoders have a relatively simple architecture with linear transformations,
which may not be expressive enough to capture complex, non-linear relationships.

39
 Fig. 22 Reconstructed fluid flow Fig. 23 Reconstructed fluid flow
snapshot using linear autoencoder snapshot using U-Net CNN

6.1.6 Using U-Net CNN

Since this model utilises the reconstructed image produced by EPOD as input, it can
effectively capture the spatial relations and produce a better output. The reconstructed
flow snapshot using U-Net CNN is shown in Fig. 22 above.

6.1.6 Using CNN-VAE

Although VAE is similar to the linear autoencoders approach, the reconstruction is
better due to their ability to learn a rich, nonlinear latent representation of the input
data.

40
 Fig. 24 Reconstructed fluid flow Fig. 25 Reconstructed fluid flow
snapshot using CNN-VAE snapshot using GAN

6.1.7 Using GAN

The GAN model is better suited to learning complex spatial-temporal correlations.
The reconstructed flow snapshot using GAN is shown in Fig. 24 above.

6.2 Root mean square error (RMSE)

Table 1
Technique/method employed RMSE value
Extended proper orthogonal decomposition 4.197
Artificial neural network 11.521
Convolutional neural network 5.043
Linear autoencoder 8.423
U-Net CNN 1.893
Generative adversarial network 2.037
CNN-VAE 2.178

41
 Chapter 7
CONCLUSIONS

7.1 Extended proper orthogonal decomposition (EPOD)

Four modes were considered while reconstructing the flow in both the time-resolved
and non-time-resolved cases. Around 98% of energy is reconstructed, as observed in
the graph.

1. The reconstruction error decreases overall as we increase the number of virtual

modes (deltr). In our case, we considered deltr = 23 for best results.

2. Reconstruction RMSE (root mean square error) per snapshot is found to vary
between 0 and 0.03

3. For this project, we have taken deltr = 23, number of modes = 4, time-resolved
snapshots = 9600 and non-time-resolved snapshots = 400.

4. RMSE (root mean square error) of the reconstruction is best found to be 4.197

7.2 Commentary on other models

ANNs and CNNs lack inherent temporal modelling: While effective for spatial
features, ANNs and CNNs struggle to capture the order and evolution of information
in time-series data like fluid flow. This limitation hinders their ability to reconstruct
flow fields that rely heavily on temporal dependencies accurately.

Autoencoders require tailored architectures for time-resolved data: Standard

autoencoders primarily focus on reconstructing static data. Reconstructing
time-resolved flow data directly with a basic autoencoder is inefficient. Future work
should explore modifications like recurrent connections or combining autoencoders
with LSTMs for improved handling of temporal sequences.

GAN, U-Net CNN, and CNN-VAE are the best-performing models: U-Net
achieved the best reconstruction accuracy (MSE = 3.584). This suggests its potential
for fluid flow reconstruction tasks.

42
 Chapter 8
FUTURE PROSPECTS

Combining Convolutional Neural Networks (CNNs) and Long Short-Term Memory

(LSTM) networks might offer a powerful approach for reconstructing fluid flow
using previous snapshots as input.

8.1 Combining CNN and LSTM networks

Shown below is a schematic that shows how it works.

Fig. 26 Schematic representation of CNN-LSTM hybrid network

8.1.1 Capturing spatial features with CNN

CNNs excel at extracting spatial features from data. In fluid flow reconstruction, the
input would be previous snapshots of the flow field, typically represented as 2D or
3D grids of velocity or pressure values. The CNN processes these snapshots,

43
identifying patterns and relationships between neighbouring data points. This allows
the network to learn the underlying spatial structure of the flow.

8.1.2 Handling temporal dependencies with LSTM

LSTMs are adept at capturing temporal dependencies in sequential data. In this case,
the sequence would be the series of previous snapshots. The LSTM processes each
snapshot while considering the information learned from the previous ones. This
enables the network to understand how the flow evolves over time.

By combining the CNN's ability to learn spatial features with the LSTM's capability
for handling temporal dependencies, the CNN-LSTM model can effectively
reconstruct the current state of the flow field based on the history provided by the
previous snapshots.

8.1.3 Benefits of this approach

This approach offers several advantages over traditional methods, such as:

1. Learning complex flow dynamics from data is possible without explicit

physical modelling.
2. Reconstructing high-fidelity flow fields is even with limited sensor data.
3. Reducing computational cost compared to full simulations for specific
scenarios.

8.1.4 Challenges
Training CNN-LSTM models for fluid flow reconstruction can be challenging due to
the following:

1. A large amount of data is required for effective training.

2. There is a need for careful selection and pre-processing of the input data.
3. The complexity of hyperparameter tuning for optimal performance.

Overall, CNN-LSTM offers a promising technique for fluid flow reconstruction using
previous snapshots. With continued research and development, this approach is
expected to become even more powerful and versatile in various applications.

44
 REFERENCES

1. Bo Liu, Jiupeng Tang, Haibo Huang, and Xi-Yun Lu (2020) Deep learning
methods for super-resolution reconstruction of turbulent flows:
https://pubs.aip.org/aip/pof/article/32/2/025105/1060593/Deep-learning-metho
ds-for-super-resolution

2. Weisheng Dong, Lei Zhang, Member, IEEE, Guangming Shi, Senior Member,
IEEE, and Xiaolin Wu, Fellow, IEEE (2020) Image Deblurring and
Super-Resolution by Adaptive Sparse Domain Selection and Adaptive
Regularization: https://ieeexplore.ieee.org/document/5701777

3. “Introduction to Autoencoders” by Abhijit Roy in Towards Data Science:

https://towardsdatascience.com/introduction-to-autoencoders-7a47cf4ef14b

4. Olaf Ronneberger, Philipp Fischer, and Thomas Brox (2015) U-Net:

Convolutional Networks for Biomedical Image Segmentation:
https://arxiv.org/abs/1505.04597

5. Junwei Chen, Marco Raiola, Stefano Discetti (2022) Pressure from data-driven
estimation of velocity fields using snapshot PIV and fast probes:
https://www.sciencedirect.com/science/article/pii/S0894177722000498#:~:text
=TR%20pressure%20fields%20are%20obtained,the%20pressure%20gradient
%20is%20implemented.

6. Singular Value Decomposition:

https://www.youtube.com/watch?v=nbBvuuNVfco

7. N. Benjamin Erichson, Lionel Mathelin, Zhewei Yao, Steven L. Brunton,

Michael W. Mahoney and J. Nathan Kutz (2020) Shallow neural networks for
fluid flow reconstruction with limited sensors:
https://royalsocietypublishing.org/doi/10.1098/rspa.2020.0097

8. Pierre Dubois a, Thomas Gomez a, Laurent Planckaert a, Laurent Perret (2022)

Machine learning for fluid flow reconstruction from limited measurements:
https://www.sciencedirect.com/science/article/pii/S0021999121006288

9. Fig 1: David W. Ashmore et. al (2022) Proper orthogonal decomposition of ice

velocity identifies drivers of flow variability at Sermeq Kujalleq (Jakobshavn
Isbræ): https://tc.copernicus.org/articles/16/219/2022/

45
10.Hamed Alqahtani, Manolya Kavakli, Dr. Gulshan Kumar Ahuja (2019):
https://www.researchgate.net/figure/The-general-architecture-of-GAN_fig5_33
8050169

46
MIN-400B Report (Final Evaluation).pdf
ORIGINALITY REPORT

13 %
SIMILARITY INDEX
7%
INTERNET SOURCES
8%
PUBLICATIONS
7%
STUDENT PAPERS

PRIMARY SOURCES

1
Submitted to Indian Institute of Technology
Roorkee
1%
Student Paper

2
e-archivo.uc3m.es
Internet Source 1%
3
Submitted to City University of Hong Kong
Student Paper 1%
4
Meenal R. Kale, Deepa -, Anil Kumar N, N.
Lakshmipathi Anantha, Vuda Sreenivasa Rao,
<1 %
Sanjiv Rao Godla, E. Thenmozhi. "Enhancing
Cryptojacking Detection Through Hybrid Black
Widow Optimization and Generative
Adversarial Networks", International Journal
of Advanced Computer Science and
Applications, 2024
Publication

5
export.arxiv.org
Internet Source <1 %
6
uia.brage.unit.no
Internet Source <1 %
7
www.journal.esrgroups.org
Internet Source <1 %
8
Junwei Chen, Marco Raiola, Stefano Discetti.
"Pressure from data-driven estimation of
<1 %
velocity fields using snapshot PIV and fast
probes", Experimental Thermal and Fluid
Science, 2022
Publication

9
Submitted to University of Oklahoma
Student Paper <1 %
10
Submitted to University of South Australia
Student Paper <1 %
11
Submitted to University of Lincoln
Student Paper <1 %
12
Santanu Pattanayak. "Pro Deep Learning with
TensorFlow", Springer Science and Business
<1 %
Media LLC, 2017
Publication

13
Fahdah A Almarshad, Ghada Abdalaziz
Gashgari, Abdullah I. A. Alzahrani.
<1 %
"Generative Adversarial Networks-Based
Novel Approach for Fraud Detection for the
European Cardholders 2013 Dataset", IEEE
Access, 2023
Publication
14
Submitted to Harrisburg University of Science
and Technology
<1 %
Student Paper

15
ruor.uottawa.ca
Internet Source <1 %
16
Submitted to Liverpool John Moores
University
<1 %
Student Paper

17
Submitted to University of Edinburgh
Student Paper <1 %
18
Submitted to CSU, San Diego State University
Student Paper <1 %
19
Submitted to SASTRA University
Student Paper <1 %
20
espace.library.uq.edu.au
Internet Source <1 %
21
Submitted to University of Bath
Student Paper <1 %
22
Submitted to University of Leicester
Student Paper <1 %
23
Submitted to Manchester Metropolitan
University
<1 %
Student Paper

24
link.springer.com
Internet Source <1 %
 <1 %
25
mafiadoc.com
Internet Source

26
Ayesha Rafique, Tauseef Iftikhar, Nazar Khan.
"Adversarial Placement Vector Learning",
<1 %
2019 2nd International Conference on
Advancements in Computational Sciences
(ICACS), 2019
Publication

27
Pawar, Suraj. "Physics-Guided Machine
Learning for Turbulence Closure and
<1 %
Reduced-Order Modeling", Oklahoma State
University, 2023
Publication

28
"Intelligent Computing and Networking",
Springer Science and Business Media LLC,
<1 %
2023
Publication

29
arxiv.org
Internet Source <1 %
30
Submitted to University College London
Student Paper <1 %
31
Submitted to University of Carthage
Student Paper <1 %
32
aaltodoc.aalto.fi
Internet Source <1 %
33
www.sandiego.gov
 <1 %
Internet Source

34
Submitted to Bilkent University
Student Paper <1 %
35
Submitted to Institute of Technology Carlow
Student Paper <1 %
36
Submitted to University of Hong Kong
Student Paper <1 %
37
Submitted to University of Surrey
Student Paper <1 %
38
philpapers.org
Internet Source <1 %
39
di.univ-blida.dz
Internet Source <1 %
40
Submitted to Ganpat University
Student Paper <1 %
41
Insu Choi, Woo Chang Kim. "A Graph-based
Approach to Predicting Downside Risks of
<1 %
Global Financial Market Indices Using
Dependencies on Inflation Rate-Adjusted
Condition", Research in International Business
and Finance, 2023
Publication

42
core-cms.prod.aop.cambridge.org
Internet Source <1 %
43
www.medrxiv.org
Internet Source <1 %
44
NONG YE, GAVRIEL SALVENDY. "Cognitive
engineering based knowledge representation
<1 %
in neural networks", Behaviour & Information
Technology, 1991
Publication

45
dspace.vutbr.cz
Internet Source <1 %
46
Submitted to Berlin School of Business and
Innovation
<1 %
Student Paper

47
Lecture Notes in Computer Science, 2015.
Publication <1 %
48
Submitted to University of Strathclyde
Student Paper <1 %
49
Yuta Kawachi, Yuma Koizumi, Noboru Harada.
"Complementary Set Variational Autoencoder
<1 %
for Supervised Anomaly Detection", 2018 IEEE
International Conference on Acoustics,
Speech and Signal Processing (ICASSP), 2018
Publication

50
Zhengguang Liu, Huang Chen, Zhengyong
Ren, Jingtian Tang, Zhimin Xu, Yuanpeng
<1 %
Chen, Xu Liu. "Deep learning audio
magnetotellurics inversion using residual-
 based deep convolution neural network",
Journal of Applied Geophysics, 2021
Publication

51
d197for5662m48.cloudfront.net
Internet Source <1 %
52
github.com
Internet Source <1 %
53
hal.archives-ouvertes.fr
Internet Source <1 %
54
pdffox.com
Internet Source <1 %
55
tnsroindia.org.in
Internet Source <1 %
56
"Image Analysis and Processing – ICIAP
2019", Springer Science and Business Media
<1 %
LLC, 2019
Publication

57
Santi Seguí, Michal Drozdzal, Guillem Pascual,
Petia Radeva, Carolina Malagelada, Fernando
<1 %
Azpiroz, Jordi Vitrià. "Generic feature learning
for wireless capsule endoscopy analysis",
Computers in Biology and Medicine, 2016
Publication

58
Shah, Prasham. "Design Space Exploration of
Convolutional Neural Networks for Image
<1 %
Classification.", Purdue University, 2023
Publication