Vargas 1

ECE 3101L

Prof. Andrew Pagnon

Lab Report #4

Image Processing

December 9, 2024

Cristian Vargas

(Lab Partner: Aaron Tran)

 Vargas 2

Table of Contents:

Executive Summary………………………………………………………………………………3

Theory/Background………………………………………………………………………………4

 Image Resizing……………………………………………….……….…….……………4

 Filtering Images……………………………..….………………………………..………6

 Image Equalization………………..…………………………..…...…………………….9

 Two dimensional Discrete Fourier Transform…………………………………………..10

 Gaussian Filters………………………………………………………………………….12

Lab/Data Anlysis…………………………………………………………….…………………13

 Image Resizing…………………………………………………………………………13

 Filtering Images…………………….………………………………………………..…14

 Image Equaliztion………………....…….………………………………………………15

 Two Dimensional Discrete Fourier Transformation…………………………………….16

 Gaussian Filters………………………………………………………………………….17

Error Analysis……………………………………………………………………………………19
 Vargas 3

Executive Summary:

The purpose of this report is to understand, test, and conclude our findings relating to the various

image processing methods (image filtering, image equalization, gaussian filters, etc.). This is

done by determining the theory of these processing methods, gathering data using these methods,

and concluding our findings.

 Vargas 4

Theory Background:

Image Resizing Theory:

To understanding Image Resizing, we must first make sure we understand how images

are interpreted in the digital sense. Digital images are comprised of various pixels, which

comprise the entire image that we see on a screen. The image has some resolution (1000 x 1000

for example) and can be scaled up or down by using various mathematical models.

If we intend to make an image smaller, we can scale and image down from 1000 x 1000

pixels to 500 x 500 pixels, for example. We can characterize the proportional shrinking of an

image by using the expression:

n
1
=proportion of pixels ¿
2

Using the above expression, we can determine the proportional amount of pixels left in

an image by using the amount of times we desire to reduce the image. If we desire to reduce the

1
image by 3 times, we can say that the image will have roughly of the original pixels. This also
8

7
means that of the original data is lost however. This will be noticed when the image is scaled
8

back up to its original size after the shrinking process.

If we desire to enlarge an image, we can use the following expression:

n
2 =amount of pixels added

Although we are adding more pixels to the image, we are not inherently increasing the

quality of the image, as the new pixels will be procedurally determined by the surrounding pixels

and the interpolation algorithms that are employed.

 Vargas 5

Some of the most common interpolation methods are as follows:

Nearest interpolation, which determines the value of the newest pixel by selecting a value

to copy based on what the nearest pixel is to it:

Bilinear interpolation, which determines the value of the newest pixel based on a

weighted average in a 2x2 neighborhood:

Bicubic interpolation, which determines the value of the newest pixel based on a

weighted average in a 3x3 neighborhood:

 Vargas 6

We can gather that the larger the pixel neighborhood, the smoother or softer the image is.

This is important in terms of image anti-aliasing which is intended to reduce the over sharpness

of an image in order to make it seem more like a natural image.

Filtering Images:

Just like we can apply various filters to analogue signals in a circuit, we can also apply

similar filters to images in the digital sense. In order to interpret how low-pass and high-pass

filters work with digital images, we must first understand how they work with a square wave.

Using a Low-Pass filter, we can see that the signal is more likely to pass the portion of the square

wave signal that has a lower frequency, or lower tendency to change. This is reflected in how the

output signal largely reflects the DC portion of the square wave.

 Vargas 7

When applying a High-Pass filter to a square wave signal, we can see that the portion of

the filter which is passed is the portion of the square signal which experiences the highest rate of

change (the rising and falling edge of the square signal).

In term so images and applying these filter to digital signals, we can associate the change

in the values relative to the surrounding pixel values. If a pixel is very similar to the neighboring

pixels then it can be characterized by a low frequency signal. If the pixel is far more different

relative to the surrounding pixels, then we can characterize it as a high frequency signal. The

high and low pass filters will remove each of these signals accordingly.

Filters can be done in MATLAB through the use of matrices to represent the output of the

filter relative to the center pixel and the surrounding pixel values.

In this matrix, we can see that the center pixel is passed while the surrounding pixels are

completely attenuated. This filter would act as a buffer, as the center pixel value is not changed at

all and the surrounding pixel values are not considered in determining the new center pixel value.
 Vargas 8

This filter would act as a Low Pass filter as the filter weighs each of the surrounding

pixels equally when determining the new center pixel value. This filter would smooth the digital

image and cause anti-aliasing.

This filter weighs the side values far more than the corner values. Although this still acts

like a Low Pass filter, this would smooth the image significantly less as the weight for the center

pixel is far less distributed.

The High Pass Laplacian filter amplifies the contrast between the center pixel and the

surrounding pixels. This would cause the output image to show more of the contrasting pixel

values and introduce a significant amount of noise to the final image.

 Vargas 9

This matrix can be used to shift the entire image to the left, using the pixel value which is

directly to the right of the center pixel. When this is repeated multiple times, this can cause an

extension type effect on the right side of the image, as there is no pixels to the right of the right

edge, thus it copies the pre-existing pixel values.

Image Equalization Theory

We can equalize an image by removing pixels that are disproportionately represented in

the image. The intention of doing this is the balance the image such that it isn’t overly bright or

overly dark. This is done by considering a cumulative sum histogram. A cumulative sum

histogram depicts a summation of each of the histogram values, but relative to the complete

number of data points in the image.

 Vargas 10

Two Dimensional Discrete Fourier Transform Theory Overview

Images are not just interpreted in the spatial domain, but also the frequency domain,

which is used to understand the relative changes of pixel values. This is often used by computers

in order to interpret things such as image processing and image recognition.

In the above image, we can see that as the number of lines is increased in the above

images, the frequency of the sine wave under them increases. This is showing how the frequent

change from black to white and vise versa reflects high frequency within the image.

The orientation of the lines also have an influence on the depiction of an image in the

frequency domain.

Using the line example we can expect that the resulting image in the frequency domain

will reflect in an image which is in a perpendicular orientation to the lines in the spatial domain.
 Vargas 11

From the above image, we can see that the vertical lines produce simple dots in the

frequency domain, but as the orientation of the lines is rotated or the frequency is increased, the

dots are at different angles or distances from the origin of the picture.

As we can see from the above image of a man taking a photograph, the frequency domain

image representation of the image shows lines which are perpendicular to the major lines of

contrast in the spatial domain image.

 Vargas 12

Gaussian Filters

The sigma parameter in the Gaussian function determines the standard deviation of the

Gaussian distribution and directly affects the filter’s characteristics. A gaussian low pass filter

with a larger sigma component will introduce a stronger smoothing effect to the image, applying

a stronger blur. One with a smaller sigma will cause an inverse effect and lead to a milder

blurring. A gaussian high pass filter with a larger sigma value will filter out high contrast and

sharper detail in an image, while a smaller sigma value will amplify less subtle components of

the image. The characteristics will be shown later in the report.

 Vargas 13

Lab:

Image Resizing Visualization

We will first see how the resizing algorithm alters the original image of the flower:

Finally, we can attempt to reverse the shrinking process by resizing the reduced image

using the three previously discussed methods: nearest, bilinear, and bicubic. From a visual

perspective, we observe a notable enhancement in image quality as we employ more advanced

techniques. This improvement arises from the increasing number of pixels referenced when

generating new pixels during the resizing process. As a result, the contrast between individual

pixels becomes less pronounced, producing a "smoother" and more refined image, particularly

with bicubic resizing. Conversely, the nearest neighbor method results in the most visibly

pixelated and coarse output among the three approaches.

 Vargas 14

Filtering Images

We can see that the Low Pass filter greatly smooths the originally grainy image, while the

High Pass filter greatly introduces noise. This confirms our prior understanding of how HPF and

LPF influence digital images.

We can also test the effects of the left shift matrix on the crane image.

Looking at the image, we can confirm our prior understanding of the left shift matrix to

also be true. We see not only was the image significantly shifted to the left, but we see an

extension of the right most pixels in the far left of the image.
 Vargas 15

Image Equalization

Using the crane image, we can test the effects of image equalization on digital images.

First considering the histogram of the original image, we can see a clear density of number of

pixels in the 75 to 100 range. We will note this for after the image equalization. Also, the

cumulative sum of histogram shows that there is a considerable increase in pixel density in the

middle and upper portion of the black and white color spectrum. This will be significantly

different after the equalization.

Considering the equalized image, we can see that the image has a more even distribution

of blacks and whites in it. Considering the histogram, we can see that the region that was

significantly dense has been extended laterally, showing that the region with the highest pixel

count now ranges from 75 to 150 on the black and white spectrum. Finally, the cumulative sum

of histogram is a completely linear line. This supports that the image is indeed equalized such

that the cumulative sum of histogram function increases at a constant rate throughout the image.
 Vargas 16

Two Dimensional Discrete Fourier Transform Visualization

When considering the frequency domain interpretation of the crane images, we will

mainly look at the FFT log scale interpretations of the two images. This is largely because there

isn’t much information to gather from the “FFT of Image” and “FFT center origin” images.

Looking at the FFT log scale image for the top crane picture, we see that there is a large

density of white pixels at the origin of the image, but a gradient to black pixels as we move away

from the center of the image. This is largely because this image of the crane is the less grainy

image, meaning that there is less noise in the image. Due to the lower level of noise in the image,

the frequency domain interpretation is populated closer to the origin but not as much on the outer

limits.

Looking at the bottom image of the crane however, we can see that the FFT log scale

image is comprised of allot of white pixels throughout the image. This is because this image of
 Vargas 17

the crane is the more grainy image, so the noise throughout the image is coming up on the FFT

log scale capture.

Gaussian Filters

First considering a sigma value of 50, we can see how this will affect the various filtering

methods of the crane image:

First considering the low pass filtering of the top, less noisy, crane image we can see the

image has been considerably smoothed. Next, looking at the high pass filter image, we can see

that the regions of the image that are still left are subtle outlines of the various edges in the

image. Considering that the low pass filter and the high pass filter removed complementary

portions of the image, the sum of the two images resulted in the recreation of the original image.

Now considering the bottom set of images, we notice that we will be passing the grainy

image of the crane through the gaussian filters. First, through the low pass filter, we notice the

subtle smoothing of the grainy image, but the more interesting result is when we take the high
 Vargas 18

pass filter over the grainy crane image. When passing the image through the filter, as the purpose

of the high pass filter is to show the high contrast portions of the image, we see that the image is

practically unintelligible, due to the high amount of noise in the image. Similarly to the previous

image however, the original image can be reconstructed using the sum of the two complementary

output images.

Let us now consider how the Gaussian filter acts differently when the sigma value is

reduced to 2:

Considering the non-grainy crane image first, we see that when passed through the low

pass filter, the image is aggressively smoothed, making the subject basically undistinguishable.

The high pass filter however, is so gentle that the only a mild reduction in brightness is done in

the more low contrast portions of the image. Both of these characteristics are notices also in the

grainy crane image, with the exception of more noise is passed in the high pass filter of the

grainy crane image.

 Vargas 19

Error Analysis:

In conclusion, the experiment was a complete success. We were able to properly

understand the properties of image processing and how images can be altered using various

filters and interpretations in the frequency domain. The lab provided valuable insights into the

application of Fourier transforms and how spatial and frequency domains are interconnected in

image analysis. Additionally, the results demonstrated how specific filters can enhance, suppress,

or modify image features, showcasing their practical applications in areas such as noise

reduction, edge detection, and equalization. Overall, this experiment deepened our understanding

of image processing techniques and their relevance in real-world scenarios. The topics covered in

this lab report could be applied in projects relating to object recognition in image processing,

facial recognition, and much more.