A

REPORT

ON

IMAGE PROCESSING AND ANALYSIS

Name of the Student: J Vedarutvija ID.No. 19STUCHH010281

AT

National Informatics Centre, Hyderabad

Faculty of Science & Technology,

The ICFAI Foundation for Higher Education, Hyderabad

2022

1
 A

REPORT

ON

IMAGE PROCESSING AND ANALYSIS

Name of the Student: J Vedarutvija ID.No. 19STUCHH010281

Under the guidance of

Ms. Radha and Ms. Annapurna

A Project report submitted for academic fulfillment for Internship

National Informatics Centre, Hyderabad

2022

2
 DECLARATION

I, Joopally Vedarutvija hereby declare that my Project Report entitled “IMAGE PROCESSING
AND ANALYSIS” is a bonafide record of the work which is uniquely prepared by me on the basis
of my own work that I have submitted to ICFAI Foundation for Higher Education in academic
fulfillment of Internship – II at IP station National Informatics Centre, Hyderabad.

Signature of the Student

3
 Faculty of Science & Technology, IFHE University

Station: National Informatics Centre Centre: Hyderabad

Duration: 6 months Date of Start: 01-July-2022

Date of Submission: 16-December-2022

Title of the Project: Image Processing and Analysis

Name of the student: Joopally Vedarutvija

Name of the Company’s mentors: Ms. Annapurna, Technical Director (Scientist-E),

Key Words:
• Computer Vision
• Image Processing
• Pillow
• Open CV
• Tensorflow
• Pytorch
Project Area: Image Comparison and Analysis

Signature of Company Mentor Signature of IP Mentor

4
Acknowledgment
I’m glad to be working on this research project i.e, Image Processing and Analysis under the
guidance of Ms. Annapurna madam and under the supervision of Ms. Radha madam and Dr.
Avinash Malladi sir

Thanks to Avinash Sir for his kind and understanding nature, giving us the freedom to work for
NIC without any evaluation pressure on us. Making sure that we are comfortable in working and
safe at the company. I'd want to express my gratitude to Avinash Malladi sir for continuously
taking updates about our work and if we are facing any difficulties.

Thanks to our point of contact Radha mam from the company for giving me this opportunity to
work on the Artificial Intelligence Project under Annapurna mam. A special thanks for giving us
the freedom in working on the projects that interests us with good consent. It gives pleasure in
discussing environmental projects that she assigns tasks, without her support, it would have been
difficult to work on projects other than AI at NIC.

Special thanks to my guide and mentor Annapurna mam for showing immense support in the
completion of this project with a great understanding and kind nature. It brings immense pleasure in
working under my mentor who is always available for me for any kind of help within or outside the
project work. The step-by-step aligned tasks that madam assigns us gives a great clarity in
understanding the concepts deeper and better. The interlinked technical tasks bring us hands-on
experience that helps me in my future research projects. It would have not been possible without
her guidance and support.

Thanks to IcfaiTech for providing such an amazing opportunity to work at the National Informatics
Centre for a duration of 6 months to foster our practical knowledge and skills.

5
Abstract
This research work aims to predict which image format and library/framework are better for
performing image processing tasks. The image formats considered in our project work are JPEG or
JPG, BMP, PNG, TIFF, and PNG.
We started by comparing two same images of different resolutions, different sizes, and different
formats (JPG, PNG, BMP, GIF and TIFF ) to find out which format is the best and how well python
libraries or frameworks work in identifying the differences in the images before and after resizing
them.
Next, we dived a little deeper and integrated the task with a java program to get output from the
bulk of images. First, uploaded the images along with their RGB data values in the PostgreSQL
database and retrieved them back onto the web page. Next downloaded the data from the database
as a CSV file and compared the images using python with the given input data iteratively along all
the data columns.
After we are successful in performing all the above tasks by our analysis we come to a final
conclusion that bitmap images are good for image processing and pillow library in addition with
TensorFlow framework works is better than OpenCV, and TensorFlow is one of the best for
computer vision as it is a framework and is provided with necessary tools for better processing.The
OpenCV library needs modifications to work better.

6
TABLE OF CONTENTS
1 Introduction…………………………………………………………………………………….…8
1.2 Computer Vision…………………………………………………………………………..... 11

1.3 Pixels…………………………………………………………………….…………….….….12

1.4 Image…………………………………………………………….………………………..….12

1.4.1 Raster Image…………………………………………………………………………….12

1.4.2 Vector Image……………………………………………………….……………………13

1.5 RGB values…………………………………………………………………………………….13

1.6 Image Formats…..………………………………………………………….…………………..14

2 Python Libraries and Frameworks…………………………………………………………….19

2.1 OpenCV……………………………………………………………………………………….19

2.2 Pillow………………………………………………………………………………………….20

2.3 TensorFlow…………………………………………………………………………………...21

3 Image Comparison………………………………………………………………………………22

3.1 Comparison of two different images…………………………………………………………..22

3.2 Comparison of two same images with different image properties……………………….…...24

3.2.1 Comparison of two Same images of different Resolution……………….…………..……24

3.2.2 Comparison of two Same images of different Size……. …………………..……………..25

3.2.3 Comparison of two Same images of different Formats …………………………………..26

3.3 Grayscale with Grayscale…………………………………………………………..…………..33

4 Inserting and retriving images from DB...…………………………..…………………………35

5 Comparison of frameworks/libraries for processing.…………………………………………38

8 Conclusion ……………………………………..…………………….…………………………..46

7
References……...…………………………………………………………………………………..48

Appendices……………………...………………………………………………………………….49

Glossary…………………………………………………………………………………………….50

8
Chapter 1 – Introduction to Image Processing
Image Processing is a type of signal processing that involves performing operations on images for

obtaining desired output. There are many image processing techniques we can use in predicting the

results using advanced algorithms in machine learning and deep learning. In our research, we

compared two same images of different formats, different resolutions, and different sizes and

analyzed the results. All the image file formats don’t have the same information stored, and each

file has its own applications and advantages.

Working with image processing techniques, it is important to understand the difference in the

information stored in the digital images that we process. One has to keep in mind which file format

helps in what kind of scenarios.

In our study, we used the famous library OpenCV and Pillow library to work on the images to

understand the differences in them.

As technology is growing drastically day by day, undoubtedly there is a rapid growth in the internet

for storage. Each day billions of images are clicked. Collecting and analyzing those images may

give us amazing results. There are many different applications of image processing in the areas of

space, medical, traffic, and banking, such as planet discovery, brain tumor detection, fraud

detection, and object detection etc. The objective of this research is to compare and analysis the

images and their formats then conclude the best format for performing image processing tasks. We

used open-source python libraries, OpenCV, and Pillow which are helpful in the computer vision

area.

9
1.1 Integrated Development Environment

Anaconda is a Python and R programming language distribution aimed for simplifying package

management and deployment in scientific computing (data science, machine learning applications,

large-scale data processing, predictive analytics, and so on). Data science software for Windows,

Linux, and macOS are included in it. Anaconda, Inc., was created in 2012 by Peter Wang and

Travis Oliphant, and developers maintain it. Anaconda comes with Jupyter Notebook, the software

is free to download and deploy the python codes.

A Jupyter notebook is an open source IDE (Integrated Development Environment) which is used to

work on data science tasks using languages like Python, Julia, Scala, R etc. But in our work we

only use Python programming language.

Notebooks are becoming an important element of the data science process at firms all around the

world, they are also very helpful in teaching data Science to students in universities and Schools as

it's a single page in which we can execute code, view the results, and add explanations, formulae,

and charts to make our work more clear, intelligible, repeatable, and sharable. If we want to work

with data, having a Notebook can help us get there faster and make it easier to discuss and share our

findings.

The python libraries and packages we’ll use in this project are namely:

10
 • NumPy

• Pandas

• Pillow

• Matplotlib

• OpenCV

• Scikit-learn

• Tensorflow

1.2 Computer Vision

Computer vision is the process of understanding how pictures and videos are stored and how we

may change and retrieve data from them. It is the field of Artificial Intelligence which trains to

understand and interpret the visual world. It is becoming increasingly important in self-driving cars,

robots, and picture-editing apps. Computer vision models can easily understand digital images from

cameras and videos using deep learning techniques but it is still a huge research area to train the

models to work even better. There are many different python libraries used in support of computer

vision to perform visual tasks. The next chapter clearly mentions about the python libraries and

frameworks that we used in our project.

11
1.3 Pixels

Pixels are the core elements of the images, abbreviated as px. The size of the pixel depends on how

we set the resolution, the higher the resolution, the higher the pixel values, and vice versa. Each

pixel is a sample unit of the original image, higher samples give an accurate representation of the

original image which is clear. A simple way to describe each pixel is by using a combination of the

three colors Red, Blue and Green (RGB). If all three values are 255 (infinity) that means it is a

white color and if all of them are set to 0 it shows black color. As each digit is an 8-bit digit the

values range from 0-255.

1.4 Image

An image is a numeric representation of images. A digital image is a 2D representation of an object

and is processed as signals by optical devices. The two fundamental image format types are raster

images and vector images (some formats, however, allow a mix of the two).

12
1.4.1 Raster image: Raster images contain pixel values arranged in a matrix fashion. These are the

pictures that result from scanning or photographing an item. Raster pictures are created by

combining pixels, or small dots, each of which has unique colour and tone information.

Raster pictures are resolution dependent since they are pixel-based. The number of pixels in a

picture, as well as the number of pixels displayed per inch, both determine image quality. As you

can expect, the greater the resolution and the number of pixels in an image, the higher the quality of

the image. Raster image quality decreases with resizing.

1.4.2 Vector image: Vector images are characterized as geometric objects using continuous

coordinates. The image quality of these types of images is good compared to raster images. They

are infinitely scalable as they are made of line art or paths. Examples of vector images are SVG,

PDF, and EPS and examples of raster images are JPG, PN, BMP, and GIF. The resolution of

vector images doesn’t change with resizing.

In our research, we mainly focus on raster images and work on them to find the differences and

analyze the results.

1.5 RGB Values

An RGB picture is three images layered on top of each other (a red scale image, a green scale

image, and a blue scale image), and when fed into the red, green, and blue inputs of a colour

display, it forms a colour image on the screen.

Because of the accuracy with which a real-life image can be duplicated, an RGB image is also

referred to as a genuine colour image.

Each parameter (red, green, and blue) specifies the color's intensity with a value between 0 and 255.

This means that there are 256 × 256 x 256 = 16777216 different colors.
13
For example, RGB (255, 0, 0) is rendered as red because red is set to its greatest value 255, while

the other two (green and blue) are set to 0.

Another example is RGB (0, 255, 0), which is shown as green because green is set to its maximum

value 255, while the other two (red and blue) are set to 0.

To display black, set all color parameters to 0, as seen below: RGB (0, 0, 0).

To display white, set all color parameters to 255, as seen here: RGB (255, 255, 255).

1.6 Types of Image Formats

There are different image file formats and each has its own advantages and disadvantages. In this

section, we will study each image format in detail. We have considered JPEG/JPG, PNG, BMP, and

GIF image formats for our analysis.

1.6.1 JPG

JPG (Joint Photographic Experts Group) is a standard image file format and requires less storage

than other formats. It uses a lossy compression algorithm which means there is a loss in the image

information while reducing the number of bits to represent the given information unlike PNG

(Portable Network Graphics), BMP(Bitmap Image File), and GIF. Alpha channel is absent in JPG.

JPG is similar to JPEG with an extension of .jpg.

The format is widely chosen for storing and sending photographic pictures over the internet.

Almost all operating systems now provide viewers that handle the display of JPEG pictures, which

are frequently saved with the JPG extension. The steps involved in JPG file compression is as

follows.

14
Steps involved in Compression:

Transformation: Color photos are converted from RGB to luminance/chrominance images (the

eye is sensitive to brightness rather than chrominance, therefore the chrominance component can

lose a lot of data and thus be significantly compressed).

Down Sampling: The coloured component is down sampled, not the brightness component. Down

sampling is done in a 2:1 horizontally and 1:1 vertically ratio (2h 1 V). As a result of not touching

the 'y' component, the image shrinks in size while maintaining visual quality.

Organizing in groups: Each colour component's pixels are grouped in groups of 82 pixels called

"data units." If the number of rows or columns is not a multiple of 8, the bottom row and rightmost

columns are duplicated.

Discrete Cosine Transform: Discrete Cosine Transform (DCT) is performed to each data unit to

provide an 8-map of changed components. Due to the limited accuracy of computer arithmetic,

DCT implies some information loss. This implies that even if the map is not there, there will be

some loss in image quality, but it is usually little.

Quantization: Each of the 64 altered components in the data unit is divided by a distinct value

known as the QC and then rounded to an integer. This is where information is irretrievably lost, and

high QC causes even more loss. In general, most JPEG implementations permit the use of QC

tables specified by the JPEG standard.

Encoding: Each data unit's 64 quantized transformed coefficients (now integers) are encoded using

a mix of RLE and Huffman coding.

15
Adding Header: The last step adds a header and all of the JPEG parameters that were used, and

then outputs the result.

The JPEG decoder reverses the processes to create the original picture from the compressed one.

2.2 PNG

PNG (Portable Network Graphics) pictures can be palette-based (with palettes of 24-bit RGB or 32-

bit RGBA colours), grayscale (with or without an alpha channel for transparency), or full-color

RGB or RGBA images. Because the PNG working group created the format for Internet image

transfer rather than professional-quality print graphics, non-RGB colour spaces such as CMYK are

not supported. A PNG file comprises a single picture in an extensible chunk structure, storing basic

pixels as well as additional information such as textual comments and integrity checks.PNG files

support transparent backgrounds which are preferable for graphic designs and can also be used as

frames for making a GIF image. It can be with or without an alpha channel and supports up to 16

million colors.

Medical Imaging is the main application area of PNG images because the reconstructed image after

decompression is exactly the same as the original image. The PNG files are stored with an

extension .png.

2.3 BMP

Bitmap Image Files (BMP) on the other hand contain bitmap graphics data, they are independent of

the device so they are known as Device Independent Bitmap (DIB). This independence allows the

file to be opened on different systems, including Microsoft Windows and Mac. Image data is

16
usually uncompressed or compressed with a lossless compression algorithm. It also contains an

Alpha channel like a PNG file and It occupies more space.

The BMP file format may store two-dimensional digital pictures in monochrome and colour, with

varying colour depths and optional data compression, alpha channels, and colour profiles. The BMP

file format is covered by the Windows Metafile (WMF) standard.

2.4 GIF

GIF(Graphics Interchange Format) uses a lossless compression algorithm it occupies less space and

is limited to 256 colors. The image contains multiple frames. Multiple JPG or PNG files together

make up a GIF file. Because it is animated video graphics. The format allows for up to 8 bits per

pixel for each picture, allowing a single image to refer to its own palette of up to 256 distinct

colours selected from the 24-bit RGB colour space. It also enables animations and has a palette of

up to 256 colours for each frame. Because of these palette constraints, GIF is less appropriate for

replicating colour pictures and other images with colour gradients, but it is ideal for simpler images

such as designs or logos with solid colour sections. GIF files feature a fixed-sized graphical region

that contains zero or more pictures. In the case of animated GIF, certain GIF files break the fixed-

sized graphical region or blocks into sub-images capable of operating as animated frames. The GIF

format stores bitmap data at pixel depths ranging from 1 to 8 bits. To save photos, the RGB colour

model and palette data are always utilised. A fixed-length header ("GIF87a" or "GIF89a") specifies

the start of a standard GIF file, depending on the version.

2.5 TIFF

TIFF or TIF, Tagged Image File Format, is a file format that represents raster images for use on a

range of devices that support this file format standard. It may describe bilevel, grayscale, palette-

17
color, and full-color picture data in a variety of colour spaces. It supports both lossy and lossless

compression algorithms, allowing users of the format to select between space and time. The format

is not machine dependent and is not limited by constraints such as processor, operating system, or

file systems. The tiff images are saved as .tif format. After a series of discussions with several

scanner manufacturers and software developers, Aldus Corporation invented the TIFF file format in

the fall of 1986. The fundamental goal of the TIFF file format was to create a standardized scanned

image file format for all desktop scanner suppliers. With the passage of time, the format expanded

from support for solely binary image formats to support for grayscale and colour images. Because

there were two previous draught releases, the original version of TIFF file format standards can be

classified as Reivision 3.0. In 1988, Revision 5.0 was released, which provided support for palette

colour pictures and LZW compression.

18
2 Python Libraries and Frameworks
A library is a collection of packages that perform specific operations whereas a framework contains

the basic flow and architecture of an application. The major difference between them is the

complexity. Libraries contain a number of methods that a developer can just call whenever they

write code. In our project, we used Python’s PIL and OpenCV libraries whereas Tensorflow and

Pytorch frameworks

2.1 Open CV

OpenCV is a large open-source library for computer vision, machine learning, and image processing

tasks and it currently plays a significant part in real-time projects, which is critical in today's

systems. It can analyze photos and movies to recognize items, people, and even human handwriting

i.e, character or text recognition. Python can process the OpenCV array structure for analysis when

combined with other libraries such as NumPy. We can employ vector space and execute

mathematical operations on these characteristics to identify visual patterns and their different

features.

OpenCV by default uses integer accurate decompression method of the JPEG image. It adds half

19
pixel correction to the image while resizing and uses bilinear transformation technique. OpenCV

reads an image in BGR format by default. It is natively written in C++.

OpenCV is used to tackle a variety of problems, some of which are mentioned here.

• Automated inspection and surveillance

• number of people – count (foot traffic in a mall, etc)

• Vehicle counting on highways along with their speeds

• Interactive art installations

• Anomaly (defect) detection in the manufacturing process (the odd defective products)

• Street view image stitching

• Video/image search and retrieval

• Robot and driver-less car navigation and control

• object recognition

• Medical image analysis

• Movies – 3D structure from motion

• TV Channels advertisement recognistion

• Face recognition

2.2 Pillow

The Python Imaging Library (PIL) extends the Python interpreter's image-processing capabilities. It

includes lightweight image processing tools that help with picture editing, creation, and storing. This

library supports a wide range of file formats, has an efficient internal representation, and has robust

image processing features.

20
The core image library was created to provide quick access to data contained in a few fundamental

pixel formats which is a good starting point for broad image processing tasks.

Pillow is compatible with a wide range of image file types, including BMP, PNG, JPEG, and TIFF.

By providing new file decoders, the library supports the addition of support for future formats.

Pillow neither use half pixel correction method nor bilinear transformation technique for resizing of

the images. Pillow reads images in RGB format only. Pillow reads all the channels of the image by

default.

2.3 Tensorflow

TensorFlow is a framework developed by Google researchers to run machine learning deep learning

and other predictive and statistical workloads. It provides stable C++ and Python APIs, as well as

non-guaranteed backward-compatible API for other programming languages. Tensorflow by default

uses Discrete Cosine Transform.

There are two steps involved in opening a JPEG image using Tensorflow they are

Step 1: First step is to open the picture u.sing Tensorflow,

Step 2: Second step is to perform JPEG decoding because JPEG is a compressed image.

It can be referred from the introduction chapter in JPEG image format section.

Tensorflow is greatly used for deep learning and Machine learning models. Tensorflow works with

tensors. And for image resizing method expects a 4D tensor and returns a 4D tensor output only. To

make Tensorflow Consistant with OpenCV, we have to expand the image from three to four, then

resize the image and squeeze the dimentions back to three. As the image will still be a tensor we

should create and run a TensorFlow session to get back our resized image in NumPy format as

output.

21
3 Image Comparison
Image differencing is a method used in image processing to discover differences between pictures.

The difference between two photos is computed by calculating the difference between each pixel in

each image and then constructing an image based on the result. To use this approach, the two

pictures must first be aligned such that their corresponding points coincide, and their photometric

values must be made compatible, either by meticulous calibration or through post-processing (using

colour mapping). The amount of pre-processing required before differencing varies depending on

the type of picture.

3.1 Comparison of two different images

Comparing two completely different images( a rose with a sunflower) using Open CV and Pillow:

22
Graphical image output using OpenCV library

Graphical image output using Pillow library

23
3.2 Comparison of two same images with different image properties

While comparison of images there can be huge varieties of image properties that can make an

image different, we have chosen the most important and effective image properties to analyze the

differences in the images

3.2.1 Comparison of two same images of different resolution

Resolution is crucial in image processing to analyze and study differences between two images. It

also helps us a lot in object detection.

Changing the resolution (DPI – Dots per Inch ) of the image outputs both images as different with

OpenCV. The same is the case with Pillow library. When an image with low quality is compared to

an image of high quality, the PPI (Pixel Per Inch) varies in both images which makes a difference

in the quality of the image. The more the value higher the quality and vice versa. When we decrease

the quality and compare with high-quality image, the output we obtained is as shown using a pillow

library.

24
But when the same task is performed using OpenCV, we hardly see any difference in the graphical

image using our naked eye though the pixel values are different we see just a black image.

3.2.2 Comparison of two same images of different Size

Comparing two images of different sizes gave us a completely black image as the output using the

OpenCV library. This implies the two images are equal, but when we compared the pixel values of

the two images, the values came out to be different. We also compared the images using the Pillow

library, and surprisingly we found the graphical image difference of the two images in not

completely black. The output image is as follows:

25
3.2.3 Comparison of two same images of different formats

As we have studied in Chapter 1, there are different image file formats and each has its own

advantages and disadvantages. In this section, we will compare the differences in the image formats

and analyze them. We have considered JPG as a standard image and compared with all the other

formats like, PNG, BMP, and GIF. We also compared other possible formats along the process.

Comparison of One format with the other is as followed:

3.2.3.1 PNG with JPG

When the JPG image is compared with the PNG image we find a few of the pixel values different,

it can be due to the background that we get convinced when we visualize having both of the

pictures aside. This is because PNG files support transparent backgrounds which are preferable for

graphic designs and can also be used as frames for making a GIF image. It can be with or without

an alpha channel and supports up to 16 million colors.

26
Medical Imaging is the main application area of PNG images because the reconstructed image after

decompression is exactly the same as the original image.

The graphical image difference between both JPG and PNG using OpenCV is shown below

We get an error comparing both images using Pillow libray because PNG has an extra alpha

channel in it the reason for not getting an error using OpenCV is that it only compares three

channels from both the images. Inorder to compare the images using Pillow library, we have to

convert both images in RGB form and then perform comparison on the images. The graphical

image output after conversion is as shown below

27
 The difference is seen even more clearly using Pillow library.

3.2.3.2 BMP with JPG

When BMP format is compared with JPG in OpenCV we get the output image which is completely

black, this happens when both the image’s pixel values get matched and when subtracted we get 0

values indicating a black color this shows the images are similar. Hence the graphical image output

28
is completely black image using OpenCV library. But when the same task is performed using the

Pillow library, the pixel-wise comparison of two images gave false as the output. Surprisingly the

graphical image output of the difference is not a completely black image. We observed some light

blue sparks on the graphical image output.

Then after our analysis, we found that the OpenCV library only compared 3 channels from the

BMP file whereas the pillow library also compared the fourth channel i.e., alpha(same in the case

with PNG as well). Hence the output image is as follows with light blue lines on the image output

as highlighted.

29
3.2.3.3 GIF with JPG

Multiple JPG or PNG files together make up a GIF file. Because it is animated video graphics.

Hence the comparison of a JPG or PNG file format with a GIF file is unveiled. It is not possible to

compare GIF, a multi-frame image with JPG, which is a single-frame image. In order to achieve

this, the frames from the GIF image file need to be extracted to be compared with a JPG file with a

single frame iteratively. But doing this is not useful, because in this case, we will be comparing a

JPG file with a JPG file that is present in GIF. So, to accomplish this task we have converted a GIF

file to RGB format and then compared it with the RGB format of a JPG image. We got the

difference in a color map fashion. The difference can be seen below.

30
3.2.3.4 GIF with PNG

In the same way as the previous section above, when a GIF image is compared with a PNG file, the

graphical image output is as shown below

3.2.3.5 TIFF with BMP

A TIFF image format is converted into BMP format and then compared with the original TIFF

image and analyzed the results

The graphical image output is shown in the next page when both images are directly compared

It can be clearly seen that there are some layer-like paint structures observed which is similar to

when we use the same color colour-pencil and highlight it dark at some places and light at other

places.

31
Image pixel values in both images are equal with no difference in their size or resolution.

Since the image pixel values are equal, we got a complete black image as the output when their

RGB images are compared which doesn’t mean both the images are same and equal.

The output when the images are compared after converting them into RGB can be seen below.

32
We got a completely black image using RGB-wise comparison of the images, so RGB comparison

doesn’t conclude the images to be the same or different, from the above research it is clear that,

though the pixel values are the same images are different.

3.3 Grayscale with Grayscale

Grayscale pictures are a subset of bilevel images. Bilevel pictures can only contain black-and-white

33
image data, but grayscale images can hold shades of grey. Grayscale photos in Baseline TIFF can

be saved as uncompressed data or compressed using the PackBits technique.

Baseline TIFF grayscale pictures can have values between 4 and 8, allowing for either 16 or 256

different shades of grey.

In order to compare a grayscale image with another grayscale image we have to first extract their

RGB values or the binary data (01) and then compare the binary values with each other to check if

they are matching. The output is a boolean variable i.e, if the images match we get a true value

output and if not we get a false value as output.

In the next section, we work with multiple bulk-class images which are stored in the database using

Java Programming language and are retrieved as a CSV file. For the same, we use a binary

comparison technique iteratively over all the cells of the column to identify the class of the image.

34
4. Inserting and Retriving Images from DB

Till 3rd chapter we understood python libraries and frameworks using the python programming

language, now we will use java programming to insert the images and retrieve them to compare

bulk images using RGB or binary comparison method

4.1 Eclipse IDE

used Integrated Development Environment (IDE) to run java programs.

Multiple languages are supported and other features may be readily combined into any of the preset

eclipse packages,. We can install the software online for free.

Above is the workplace of eclipse IDE to run any java programs so we will use this as our IDE to

connect with the database in the next section.

35
4.2 PostgreSQL
We use postgres as our backend database to store the data and work on it. PostgreSQL is a

sophisticated, open-source object-relational database system that has been actively developed for

over 35 years and has a high reputation for dependability, feature robustness, and speed.

Postgres is similar to that of SQL, where in we have to run SQL queries to interact, change or

update the data tables.

So first we have to install Tomcat Server on Eclipse IDE and connect it, then we run our program

on the webpage as it is visible to us what changes we make in the program. After retrieving the

images on to the webpage using the Java Code we need to download the CSV file which contains

the Binary Data of the stored images. The next step is the insert the CSV file back into jupyter

notebook and use a for loop to run and compare the Binary data iteratively to check if the data

matches with the given input and output the result if the match is found.

36
In the same way the Binary data is compared and the result of the image class is mapped with the id

of the image data.

37
5 Comparison of frameworks/libraries for Processing

Modern Computer Vision (CV) is a trendy topic of study that mostly includes dealing with pictures.

To do this, one must utilise a framework to open the photos and perform some processing on them.

In today's fast-paced framework development, each framework has its own manner of managing

pictures, each with its own set of standards. As a result, CV solutions designed in one framework

may not operate as intended in another. It might take many days to figure out what went wrong,

which can significantly delay the project. In this chapter we have compared the frameworks with

libraries and shown the results.

We first used OpenCV library to compare with the reults with that of Tensorflow and then continue

the same process with Pillow library.

Code to open any image using OpenCV:

Here we have taken a sunflower image as input.

38
Something is odd with the figure above. It does not resemble the original. We first have to

understanf how colours in photographs are represented. In general, pictures have three colour

channels that create the colours in the pixel: red, green, and blue (RGB). The sequence of these

channels affects the colour of the pixels, as pixels always perceive the first channel as red, the

second as green, and the third as blue.

However, when the image is shown, the pixels misread the channels (i.e. pixels get confused and

interpret Blue as Red and vice-versa). That is why the picture above has the incorrect colours. So,

we transform the picture from BGR to RGB after opening it. We do this by using the OpenCV

setting COLOR BGR2RGB. The code below demonstrates how to properly open an image with

OpenCV. After understanding how colour channels function in an image, ws investigated why we

39
observed a different colour image above. When we open an image using OpenCV, it opens the

picture in the Blue, Green, and Red channels by default (BGR).

Then we opened a picture in Tensorflow and checked whether we obtain the same results. It takes

two steps to open a JPEG picture in Tensorflow. The first step is to open the picture, followed by

JPEG decoding because JPEG is a compressed image.

Code to open an image using Tensors:

40
Visually, the images read by OpenCV and Tensorflow appear to be the same. Subtracting the two

we discover a significant differences. If the two are identical, the picture should be totally black.

But we have seen a huge difference. This disparity is due to the fact that OpenCV utilises integer

correct decompression of the JPEG picture by default. TensorFlow, on the other hand, employs the

Discrete Cosine Transform as the default as mentioned in Chapter 3 about frameworks and

libraries. This method of decoding is wrong in tensorflow, we must decode it using integer-accurate

decompression to match OpenCV. Setting the option dct method='INTEGER ACCURATE' as

shown below will do this.

And then, when we compare it with the OpenCV picture, we obtain a black image. This indicates

that both our TF and OpenCV are now interpreting the picture consistently. This is an important

stage because even little variations in how we read the image might result in major differences in

how the image is processed later.

41
The result we obtained after comparing the resized pictures from OpenCV and TF.

And now we got the anticipated outcome. Finally, Tensorflow and OpenCV are continuously

reading and resizing pictures.

When moving from one framework to another, minute details like what we have seen must be

reviewed and maintained. Otherwise, they will produce different outcomes in different frameworks,

which would have a negative impact on our projects.

Differences between PIL and OpenCV are analyzed below. First, we read the image using PiL

library.

Opening an Image using Pillow:

42
The output after comparing the results of Tensorflow with the Pillow image just like we did it using

OpenCV is as shown:

This indicates that, unlike OpenCV, Pillow and Tensorflow read the image properly in RGB

format.

Then we tested if there is any difference between the JPEG picture opened by PIL and the one

opened by OpenCV

43
There is a lot of difference as we can see above. Hence the resizing mechanism of the two

frameworks differs significantly once again. Even when we opened the pictures in both frameworks

regularly, the resizing approach produced different outcomes. By default, the OpenCV resize

method employs bilinear transformation. With Tensorflow, we applied the same bilinear technique.

Nonetheless, our outcomes differed. This occurred as a result of OpenCV's usage of half-pixel

adjustments when scaling images. Tensorflow, on the other hand, does not by default. This sums up

the resizing method output differences. To address this issue, there is a parameter in the

TensorFlow bilinear resize that does half-pixel correction is shown below:

44
As a result, the desired difference in resizing is obtained. Because the PIL library does not allow

half-pixel correction, it is not enabled by default. We may assume the same findings when

evaluating the difference between PIL and TF because we know TF and OpenCV are consistent.

As a result, while utilizing the PIL framework, we should keep in mind that switching to another

framework is not an option if the task is dependent on picture pre-processing, as seen above. This

will have unanticipated consequences.

45
6 Conclusion

The images of different sizes, different resolutions, and different image formats have been

compared and analyzed. We have used two different libraries to achieve this. OpenCV (Open-

Source Computer Vision Library), developed by Intel, and Pillow (Python Imaging Library) by

Fredrik Lundh and Contributors and a Framework Tensorflow.

In conclusion, we found that the pillow library is friendlier compared to OpenCV and gives the

correct output. Because the pillow library can be used to compare images in depth. It uses all the

channels while comparing two images including the alpha channel. OpenCV is built in C and C++,

but PIL is written in Python and C, When dealing with 1000s of photos for data extraction,

processing the speed is critical, so after resizing images Pillow was faster in prosessing.

But there are some problems using OpenCV, it only compares three channels which shows the

pixelwise output is the same and it outputs both the images are completely the same, which we

concluded as wrong in our case with BMP file format. Similarly, OpenCV outputted a completely

black graphical image when two images of different sizes are compared. This can give a wrong

conclusion that both the images are the same but when the pillow library is used the difference is

shown clearly, as their pixel values are also not the same.

And From the section comparing BMP with TIFF, we concluded Bitmap image format is the best

for image processing as the pixel values are still the same the image are not the same. This

indicated the digital information captured in a Bitmap file in smaller points of bits is higher in

storing the correct image data for better processing experience.

From chapter 5 it is clear that OpenCV reads images in BGR format while TensorFlow doesn’t,

which gives us a different output than a completely black image that is expected when the results of

46
OpenCV are compared to that of Tensorflow. But when the image read using Tensorflow is

compared to the image read using Pillow library we got a completely black image as output. This

means the pillow library and TensorFlow read the images properly as the original but to do the

same in the case with OpenCV, we have to convert the BGR format to RGB and compare the

images to get a completely black image. Pillow is better at working with fewer image formats but

our research concludes Tensorflow is better to use than OpenCV. Tensorflow along with Pillow

gave better results than OpenCV. So CV solutions requires to be modified to work better for image

processing tasks. This project is successful in identifying the best format for image processing as a

Bitmap Image file and the best framework as Tensorflow.

47
References
1. Comparison and analysis of Image File Formats Gomathi Kasi Viswanathan, et al.

2. Concise Report on Image Types, Image File Format and Noise Model for Image Preprocessing,

rof. Shilpa Hiremath1, et al., 2020

3. Image file formats, Lk Tan, et al., 2006

4. Comparison and Analysis of Image Formats., Gomathi K.V.

5. Image File Formats: Past, Presnt, and Future, Richard. Wiggins III, MD, et al. 2000.

48
Appendices
Steps involved in JPEG compression……………………………………………………….15

Transformation

Down Sampling

Organizing in groups

Discreate Cosine Transform

Quantization

Encoding

Adding Header

49
Glossary
1 JPEG -Joint Photographic Experts Group………………………………….………………………………………………………………15

2 PNG – Portable Network Graphics…………………………………………………………………………………………………………..17

3 BMP – Bitmap Image File…………………………………………………………………………………………………………………………17

4 GIF – Graphic Interchange Format………………………………………………………….……………………………………………….18

5 TIFF – Tag Image File Format………………………………………………………………….……………………………………………….19

50
51