Image Quality Assessment in the Context of the Brazilian Electoral

System

Marcondes Ricarte da Silva J

unior

1 a

, Jairton da Silva Falc

ao Filho

1 b

, Zilde Souto Maior Neto

1 c

Julia Dias Tavares de Souza

Vin

ıcius Lima Ventura

and Jo

ao Marcelo Teixeira

2 d

Informatics Center, Universidade Federal de Pernambuco, Recife, Brazil

Electronics and Systems Department, Universidade Federal de Pernambuco, Recife, Brazil

Keywords:

LCD Screen, OCR, Blur, No-reference, Image Quality.

Abstract:

The Brazilian electoral system uses an electronic voting machine to increase the voting reliability. This voting

machine goes through a series of security procedures, and the one that uses the most human resources is the

integrity test. The proposed solution to optimize these resources is using a robotic arm and computer vision

methods to replace the eight persons needed to carry out the test currently. However, there is a problem

with the LCD screen in the poll worker’s terminal.There is no backlight on the LCD screen, this may cause

visual pollution on images captured by the camera, depending on the ambient lighting and camera position.

In this way, this paper proposes two methods to make it easier to choose the best images to be used in the

extraction of information process: OCR and blur analysis. We analyzed 27 images with three ambient lighting

conﬁgurations then compared our results with three no-reference image quality evaluators and research on

human perception of image quality. The OCR analysis matched very closely the human perception and the

other evaluators.

1 INTRODUCTION

The Electoral Justice is the Brazilian governmental

agency responsible for the entire electoral process

for the whole country, including all states and cities

elections. The Supreme Electoral Court (TSE) is the

highest organ on a federal level, and the headquar-

ters are based in the Brazilian capital, Brasilia. As

for the states/provinces, the Regional Electoral Courts

(TREs) are the ones responsible for the elections.

These agencies together are the authority for all steps

in the electoral process, from the electoral register to

the diplomacy of the elected, passing through the vot-

ing stage (Court, 2020).

The Brazilian electoral process started to change

in 1995 when the TSE started the genuinely Brazil-

ian electronic voting machine project. This project

replaced the slow and susceptible fraud process of

manual counting in paper ballot elections that came

with errors and widespread suspicion. Then in 1996,

https://orcid.org/0000-0003-0359-6113

https://orcid.org/0000-0001-6383-7551

https://orcid.org/0000-0001-5993-2062

https://orcid.org/0000-0001-7180-512X

in municipal elections, 57 cities had the ﬁrst contact

with the electronic voting machine, counting the vote

of more than 32 million Brazilians. Furthermore, in

the 2000 elections, Brazil started the ﬁrst fully com-

puterized election (Court, 2022c).

The electronic voting machine is a microcomputer

explicitly used in elections and is composed of two

terminals: the poll worker’s terminal and the voter

terminal. The ﬁrst consists of a numeric keyboard, a

Liquid Crystal Display (LCD), and, in some models,

a biometric reader. In this ﬁrst one, the voters’ identi-

ﬁcation and authorization to vote are carried out. The

second has a numeric keyboard and a Light Emissor

Diode (LED) screen, where the vote is privately reg-

istered (Court, 2022b).

Security tests are performed on the electronic vot-

ing machine every election year. Its primary process

is known as the “Electronic Voting Machine Integrity

Test” (“Teste de Integridade das Urnas Eletr

onicas” in

Portuguese) and aims to demonstrate the functioning

and security of the electronic voting machine. This

process begins the day before the election; a series of

local polling stations are randomly selected to have

their electronic voting machine tested. The number

selected from each state is proportional to the total

444

Silva Júnior, M., Falcão Filho, J., Neto, Z., Tavares de Souza, J., Ventura, V. and Teixeira, J.

Image Quality Assessment in the Context of the Brazilian Electoral System.

DOI: 10.5220/0011786500003417

In Proceedings of the 18th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2023) - Volume 4: VISAPP, pages

444-451

ISBN: 978-989-758-634-7; ISSN: 2184-4321

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

number of polling stations used for that voting round.

Regarding the votes that will be cast at the in-

tegrity test, the TREs provide several ballots that cor-

respond between 75% and 82% of that polling station

voters. That is because to realize a proper test is nec-

essary to account for the absentees, therefore it is not

necessary to have 100% of voters. As for the can-

didates chosen, the parties are invited to ﬁll the bal-

lots used in the audit before the election week. Also,

school kids that cannot vote yet are chosen to ﬁll bal-

lots to complete the total number of votes necessary

in the electronic voting machine. Then, on election

day, simultaneously with the actual election, the audit

begins with the veriﬁcation that there are no votes in-

side the equipment. Then, all votes obtained with the

ﬁlled ballots are entered one by one in the electronic

voting machine and into a computer at the same rate

as an auxiliary system for conﬁrmation. At the end

of the voting process, the electronic voting machine

prints a Ballot Report (BU), and the auxiliary system

also issues a report. Finally, all the data are compared,

verifying if the electronic voting machine had its nor-

mal functioning as expected (Fachin, 2021).

Approximately eight people per electronic voting

machine are involved in this process, divided in two

rounds. The ﬁrst person takes the paper ballot that

was ﬁlled out and checks its validity, the second en-

ters a valid title with the votes on the ballot in the

auxiliary system, the third qualiﬁes the voter in the

poll worker’s terminal and the fourth enters each digit

of the vote in the voter’s terminal (Court, 2022a).

To minimize human interference increasing secu-

rity and conﬁdence of the integrity test, we are de-

veloping an alternative that uses a robotic arm aided

by computer vision. In the test, the robot performs the

functions of typing the voter’s title in the poll worker’s

terminal and the votes in the voter’s terminal, remov-

ing human interference in these actions, only one per-

son is needed to change the ballot. The solution using

the robotic arm to automate the integrity test received

an innovation award.

A camera attached to the robotic arm is used to

extract, from each ballot, the title and the votes to

be typed by the robot and subsequently to view the

screen and validate what was typed at the terminals

and ensure that the information extracted from the

ballots is the same as typed on the screen, ensuring

the integrity test.

The technique used to extract the information will

be Optical Character Recognition (OCR). According

to Tekin(Tekin et al., 2011), LCD screens have some

characteristics that make them difﬁcult to read by

OCR, such as low contrast and proximity of the edge

of the numbers to the edge of the screen. In addition,

depending on the camera’s position, there may be re-

ﬂections or shadows on the screen that superimposes

the information.

The main objective of this work is to present

metrics to classify images on the screen of the poll

worker’s terminal regarding their quality for extract-

ing information and thus deﬁning the best position of

camera for extraction of information on the screen

and thus ensure a more reliable automation system

because the electronic voting machines are distributed

throughout Brazil and with that there are several light-

ing scenarios. And doing so we intend to validate the

classiﬁcation of the same images by people and three

no-reference image quality scores: BRISQUE (Mittal

et al., 2012a), PIQE (Chan and Goldsmith, 2000), and

NIQE (Mittal et al., 2012b).

The remainder of this paper is structured as fol-

lows. Section 2 list some works related to the pro-

posed solution. Section 3 describes the methodology

used to classify the images. Section 4 presents the re-

sults obtained with the analyses and metrics used. Fi-

nally, on section 5 we expose the conclusions of this

paper.

2 RELATED WORKS

In this section, we present works that discuss im-

age quality analyzer with reference and no-reference

methodology and OCR on screen.

2.1 Image Quality Analyzer

Image quality analysis is usually performed on

databases with prior human evaluation. Qin and El-

Yacoubi (Qin and El-Yacoubi, 2017) presents a Deep

Neural Network (DNN) trained to predict the qual-

ity of ﬁnger vein images is proposed. Some ap-

proaches applied to two public datasets were com-

pared, where the best results were for Patch-DNN.

Mittal, Soundararajan, and Bovik (Mittal et al.,

2012b) deal with a blind image quality analysis that

uses measurable deviations from statistical regular-

ities for natural images without training on human

opinions. The proposed index scored better than

the Popular Signal-to-Noise Ratio (PSNR) and Struc-

tural Similarity Index (SSIM) and is equivalent to

high-performance image quality analysis approaches.

Courtney (Courtney, 2021) analysis the image qual-

ity of non-referenced documents to select the best im-

age for the OCR application automatically. It used a

metric based on character feedback and checked the

parameters inﬂuencing OCR extraction, e.g., noise,

blur, contrast, and brightness, seen as the most signif-

Image Quality Assessment in the Context of the Brazilian Electoral System

445

icant inﬂuencers. The OCR solution used was Tesser-

act (Smith, 2007). The metric validation was done

with images created with inﬂuence on the parameters

mentioned compared with other indexes: BRISQUE,

NIQE, and PIQE. Furthermore, in comparison, the

suggested index could deﬁne the best image for OCR

extraction in all dataset images.

2.2 OCR

Optical character recognition is a computer vision

technique for extracting characters from images and

documents. Kastelan et al.(Kastelan et al., 2012) use

a system for OCR text extraction on the TV screen

to validate the operation of the TV by comparing

the extracted information with the expected text us-

ing Tesseract. Vellian et al. (Mariappan et al., 2011)

proposes a system for extracting characters from LED

and LCD screens in real-time, passing the informa-

tion through sound, is discussed to assist people with

visual impairments. They use adaptive binarization

and an approach to extract “blobs,” which serves to

identify the highest probability of identifying possi-

ble digits. In tests, average error rates of 1.19% and

maximum error rates of 1.52% were obtained. Tekin

et al.(Tekin et al., 2011) also use OCR for character

extraction of an LCD screen, but for multiple medical

instruments. The system works with three simultane-

ous screens, and the technology used is the LabView

Vision Assistant OCR (Kwon and Ready, 2014). The

OCR is trained with images from the LCD screen re-

peatedly until it can extract the characters accurately.

In the end, the system presented a 100% success rate

for the ten tests performed.

3 METHODOLOGY

In this section we describe all the steps used to ob-

tain the scores. We create the dataset of the terminal

screen images with the assistance of a robotic arm.

We analyzed the OCR character extraction and blur

for all images. To validate the result we compared

with three image quality evaluators no-reference and

veriﬁed the result with people’s perception with the

application of surveys.

3.1 Dataset

First, to perform the blind analysis, it was necessary

to create a dataset composed of different images that

brought to the experiment a diversity of positions and

contrasts to simulate a layman user operating the sys-

tem. Therefore, initially 9 ﬁxed positions were se-

lected for the camera focusing on the terminal screen,

being three positions with height variation: above the

screen (P1), over its center (P2) and below it (P3), in

which each one had two auxiliary positions: to the

left (P-L) and to the right (P-R). These positions are

exempliﬁed in Fig.1.

Figure 1: Illustration of positions for the camera for the ter-

minal screen. First the side view demonstrating the height

variation of the central positions, and then all positions in

the top view.

In addition to the position variation, a lighting

variation was also made in a room with little natu-

ral light irradiation. However, the place has 24 LED

tubes with 18 W of power, divided into 12 luminaries.

These are arranged in four rows with three luminaries

each. Three types of lighting were used: no lighting,

partially lit and fully on. The ﬁrst conﬁguration has

all lights off, the second with half of the lights on,

and the last with all lights on. To ensure that in each

lighting variation we would have the same relative po-

sition, we had the help of a COMAU eDO (COMAU,

2021) robotic arm. A Logitech C922 HD webcam

was attached to the robotic arm being taken to the 9

deﬁned positions, in the three different illuminations

conditions following the same order. Thus, 27 images

were obtained to be analyzed. It is important to note

that the information on the screen does not vary. Sub-

sequently, a correction was made in the images with

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

446

focus only on the poll worker’s terminal screen, iso-

lating the screen. The four points used for perspective

distortion correction were selected manually and used

a function from the OpenCV (Bradski, 2000) library

for perspective change. With the dataset ﬁnalized, the

image analysis phase began. The dataset images can

be seen in Fig.2.

In the link there is a video showing the robotic

arm passing through the positions used to obtain the

images of the poll worker’s terminal screen.

Figure 2: Dataset of images of the terminal screen.

3.2 OCR Analysis

The terminal screen has 4 lines with 40 cells per line.

For the criterion of OCR extraction each image was

cut into 160 rectangular cells, to separate the screen

image into every possible character space. Python

code was used to automate this process. Fig.3 shows

the demarcation of each cell obtained for an image of

the data.

Figure 3: Image of terminal screen with the cutout mark-

ings.

The clipping is done by ranging each line sepa-

rately. The default width for the number of cells per

line should be equal to 30 pixels due to the image

width being equal to 1200 pixels, but to try to min-

imize the errors caused by the distortion caused by

the perspective change, 2 pixels on each side were

added, totaling 34 pixels, which can be seen in the

lateral overlay of the rectangles with different colors,

in Fig. 3. There are two exceptions caused by the

side edges of the image, which makes the ﬁrst cell of

each row 30 pixels wide and the last 32 pixels wide.

The default height for the number of cells per column

should be equal to 74 pixels, since the height of the

image is 296 pixels, however, for the same reason of

the width variation, it increased 3 pixels on each side,

totaling 80 pixels, this value was used for lines 2 and

3. For the ﬁrst line the height used was 74 pixels and

for the last line 71 pixels.

The images are then pre-processed before apply-

ing the OCR. In this step, only color space transfor-

mation was used, from RGB to grayscale, and then an

adaptive thresholding from the OpenCV library was

applied, with blockSize equal to 11 and the value of

the constant that is subtracted from the average equal

to 3. This binarization technique segments the image

based on a threshold that dynamically changes over

pixel regions of the image. This sequence was used to

try to make the character clearer for the OCR and as

this is a blind analysis, the objective is not to analyze

if the OCR correctly returned the characters, but to

verify if it can identify that there are any characters.

Fig.4 shows some characters from the original image

and after pre-processing.

Figure 4: Example with 8 cells before and after pre-

processing.

With the binarized images, EasyOCR (Awal-

gaonkar et al., 2021) was used to recognize the char-

acters. This OCR was used because it needs less pre-

processing, extracting the text from the images more

directly.

After having the result of the set of cells per im-

age, an initial pattern is created as an initial reference

of the image information, being veriﬁed which char-

acter is most repeated, thus ﬁnding the mode between

a given cells in the 27 images. When there is one more

of eligible character as mode per cell, that cell in the

initial reference becomes empty. Next, a comparison

is made between the initial reference and the character

is extracted per image and thus a value is assigned to

each image, termed the initial score. This value is the

number of characters that belong to the image and the

initial reference, divided by the number of characters

contained in the initial reference, disregarding spaces.

Image Quality Assessment in the Context of the Brazilian Electoral System

447

Fig.5 exempliﬁes how the initial result was obtained.

Based on the ﬁnal score the ﬁnal reference is built.

The empty spaces that were left in the initial reference

for having more than one mode per position are ﬁlled

with the characters in which the image has the highest

initial score. Afterwards, another comparison is made

between the images and the ﬁnal reference, following

the same methodology of the initial score, to deﬁne

the ﬁnal score per image. In Fig.6 the logic that was

used is exempliﬁed.

3.3 Blur

The blur level deﬁnes how difﬁcult the objects in the

scene could be seen, eventually related to bad camera

focus. It is an important metric for evaluation because

it causes some problems for edge detection techniques

due to the absence of edges in blurred images. With

this in mind, our focus on this metric aims to deﬁne

the blur level of an image, quantifying how blurred is

an image.

Pech-Pacheco et al. (Pech-Pacheco et al., 2000)

and Bansal et al. (Bansal et al., 2016) show a proposal

for the use of the Laplacian Operator to obtain a met-

ric to evaluate blur in an image. As this operator is

often used for edge detection (Shrivakshan and Chan-

drasekar, 2012), the variance of the responses from

detected edges can provide a metric to classify an im-

age as blurred or not.

This variance deﬁnes the edge responses of the

image which the operator has applied. So, as high

it is means that there are a lot of edge responses in-

dicating that many edges are detected, and as low it

is represents detection of very little edges on the im-

age. The problem with the variance is that it is very

domain-dependent, so after analyzing how this vari-

ance performs on our dataset we decided to clamp the

values to the range [3, 30] where as low the value most

blurred is the image.

3.4 Metrics Comparison

In order to validate our approach we applied our

dataset to 4 metrics: survey, BRISQUE, PIQE, and

NIQE.

For the survey metric have been applied three sur-

veys where each have nine images of a single light

variation has been explained in Dataset subsection.

The ﬁrst with images 1-9, regarding a poorly lit room,

got 10 answers. The second covered the pictures 10-

18, regarding a partially lit room, had 11 answers.

And the third survey was answered by 20 people with

the images 19-27 taken in the fully lit room.

This metric function is to determine what a regu-

lar person perceived as a good image in a quantiﬁed

way. For that reason all three surveys had the same

ten standardized questions where which answer could

provide the image with point. These points will be

add up and then divided by the number of people on

that survey, therefor making the image ofﬁcial score.

Initially everyone was asked to rank their top 3 im-

ages in that speciﬁc poll, the ﬁrst position earn the im-

age 3 points, the second two and the third one point.

Then, so all images could be analyzed and we could

get a better take on what a person looks for a good

image we asked them to describe the images in one of

this six sentences:

• I am not able to read any character (no points);

• I am able to read some characters (2 points);

• I am able to read most of the characters and un-

derstand some words (3 points);

• I am able to read most of the characters and un-

derstand all of the words even though they are in-

complete (4 points);

• I am able to read all the characters and all words

(5 points);

Therefor all images could get up to eight points if

it got chosen as best picture and people were able to

read every character in the screen.

BRISQUE, PIQE, and NIQE were used because

they are no-reference image quality score. The ﬁrst is

Blind/Referenceless Image Spatial Quality Evaluator.

Second is Perception based Image Quality Evaluator.

And the last is Naturalness Image Quality Evaluator.

To obtain the value of these metrics the algorithms of

(Aggarwal, 2020) were used. Your scores are in the

range 0 to 100. For all, evaluators a smaller score

indicates better quality.

4 RESULTS

Here we present the results for the presented metrics.

First, we ran each metric separately on the dataset,

obtaining the scores for each image. After, we discuss

about the presented results, explaining how well each

one has behaved and discussing about them, ﬁnishing

showing our performance by making the comparison

between it and the results from the survey and some

techniques in the literature.

4.1 OCR

The ﬁnal score of each image was put into Table 1,

it can be seen that the image that scored the best is

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

448

Figure 5: Illustration the creation of the initial result.

Figure 6: Illustration the creation of the ﬁnal result.

image 24, which belongs to the subgroup of full illu-

mination and P2L position. Analyzing by subgroup,

no illumination had the worst result, with 7 images

scoring 0 and an average of 0.34 points. The best sub-

group was the one with total illumination, which has

the image with the best score and an average of 12.50

points, while partial illumination has an average of

5.90 points.

By position, P2L obtained the highest average,

with 15.10 points, and P1R had the worst average,

with 0.52 points. Among them and in descending or-

der of average, there are: P3L with 9.90 points, P3

with 8.85 points, P1L with 5.72 points, P2 with 4.69

points, P1 and P3R both with 4.17 points and P2R

with 3.13 points.

4.2 Blur

As discussed in the previous section, the Blur met-

ric provides us a variance value that depends on the

domain, where after some tests we have decided to

clamp the values in the range [3, 30]. Looking for-

ward, as our goal is to apply these values in the pro-

posed formula we have made a simple linear conver-

sion obtaining values in the range [0, 1]. Tab. 1 con-

tains the results for the blur metric for each image.

Note that the higher is the value, the sharper is the

image, so it has better quality.

As we can see, the generated results are quite rea-

sonable: for the images 1 to 9 that represents the im-

age with no luminosity we have really low values,

mainly due to the absence of edges because no fea-

tures can be really detected on the image. For the

images 10 to 18 as we have more edges recognized in

the image due to the luminosity the results increased

a lot, and for the images 19 to 27 as we have the total

luminosity the characters are most sharpness.

4.3 Comparative with Metrics and

Survey

The results obtained with the surveys are contained in

Tab. 1. Based on human perception, the image with

the best quality is 26 and the worst are 1, 2 and 3.

Table 1 contains the result for all images for OCR

analysis, blur analysis, the metrics no-reference, the

survey and voter’s title extraction, all in percent. For

OCR, blur and survey highest score indicates bet-

ter perceptual quality but for BRISQUE, PIQE and

NIQE, a smaller score is better. The best scores per

column and the number of the respective images are

in bold.

With Table 1 it can be seen that our OCR analysis

and the BRISQUE metric elected image 24 as the best

image. For our blur analysis two images are elected

the best, 19 and 23. For PIQE the best image is 2 and

for NIQE the best image is 8. According the survey

the best image is 26. For the worst images there was

Image Quality Assessment in the Context of the Brazilian Electoral System

449

Table 1: Results of OCR, BLUR, BRISQUE, PIQE, NIQE and survey.

Nº OCR BLUR BRISQUE PIQE NIQE SURVEY

1 1.56 6.92 55.66 71.53 16.09 0.00

2 0.00 1.77 64.5 14.29 12.98 0.00

3 0.00 1.82 60.68 41.4 14.36 0.00

4 0.00 1.9 64.27 80.69 14.78 17.50

5 0.00 2.66 62.06 20 16.85 37.50

6 0.00 5.18 56.85 69.03 14.5 41.25

7 0.00 5.94 57.15 68.06 14.27 36.25

8 0.00 4.28 58.79 43.78 11.91 13.75

9 1.56 2.26 61.96 32.91 12.57 10.00

10 3.12 48.61 55.79 60.15 17.89 45.50

11 1.56 16.49 63.83 69.14 15.58 19.25

12 0.00 20.59 60.6 69.52 15.33 14.75

13 4.69 33.94 59.67 49.13 13.16 71.63

14 4.69 54.78 60.17 56.02 12.69 47.75

15 12.50 40.46 58.31 52.82 13.28 71.63

16 10.94 22.53 59.46 49.08 13.19 54.50

17 14.06 9.8 59.98 44.9 13.47 77.25

18 1.56 15.34 59.59 52.93 14.12 60.25

19 12.50 100 52.59 42.33 14.36 10.63

20 10.94 29.05 61.01 54.5 14.39 2.50

21 1.56 62.81 64.97 53.08 14.07 14.38

22 4.69 59.52 58.23 48.98 15.18 65.00

23 9.38 100 55.1 45.84 13.69 43.13

24 32.81 86.43 46.26 41.71 13.59 49.38

25 18.75 46.81 60.74 56.56 14.16 52.50

26 12.50 40.3 63.92 69.13 13.29 79.38

27 9.38 47.68 62.84 62.86 14.03 70.63

Table 2: Values of Spearman correlation coefﬁcient.

SURVEY BRISQUE PIQE NIQE

OCR 0.574 -0.246 -0.094 -0.266

BLUR 0.422 -0.365 0.063 -0.018

no similar result, our OCR analysis elected image 6,

our blur analysis elected image 2, for BRISQUE is

21, for PIQE is 4, for NIQE is 10 and for survey three

images obtain zero score: 1, 2 and 3.

The best image for OCR analysis was ranked only

9th in human perception. The best images for blur

analysis were ranked 22 and 12. Images 2 and 3 were

considered the worst in the survey and in the OCR

analysis, and image 2 was also considered the worst

for blur analysis.

To verify the correlation between the metrics,

Spearman’s correlation was used. The values of the

coefﬁcients are contained in Table 2. It can be seen

that the OCR rating has a moderate correlation with

the survey while the Blur rating has a weak correla-

tion. PIQE has an almost zero correlation with both

approaches while BRISQUE and NIQE have a nega-

tive and weak correlation.

5 CONCLUSION

The work presented in this paper describes two meth-

ods used to analyze the quality of the clerk´s terminal

screen image by comparing the results with three no-

reference image quality score and a survey.

It was seen that for the best quality analysis for

LCD screen extraction the OCR analysis obtained the

best result compared to human perception and other

metrics in the literature, which is validated by having

the best correlation with human evaluation. Its use

is recommended to choose the best image to extract

information from the screen of the poll worker’s ter-

minal.

Proposition for future work would be the use of

the proposed evaluator as an OCR analysis to deﬁne

the best position for the robot used in the automation

of the electronic voting machine integrity test with

more evaluation positions and screen extraction au-

tomation.

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

450

ACKNOWLEDGEMENTS

This work has been supported by the research coop-

eration project between Softex (with funding from

the Ministry of Science, Technology and Innova-

tion—Law 8.248) and CIn-UFPE.

REFERENCES

Aggarwal, A. (2020). Buyizhiyou/nrvqa: No ref-

erence image/video quaity assessment (brisque/

niqe/piqe/diqa/deepbiq/vsfa. url= https://github.com/

buyizhiyou/NRVQA. Accessed: 2022-09-15.

Awalgaonkar, N., Bartakke, P., and Chaugule, R. (2021).

Automatic license plate recognition system using ssd.

In 2021 International Symposium of Asian Control As-

sociation on Intelligent Robotics and Industrial Au-

tomation (IRIA), pages 394–399. IEEE.

Bansal, R., Raj, G., and Choudhury, T. (2016). Blur im-

age detection using laplacian operator and open-cv.

In 2016 International Conference System Modeling &

Advancement in Research Trends (SMART), pages 63–

67. IEEE.

Bradski, G. (2000). The opencv library. In Dr. Dobb’s Jour-

nal: Software Tools for the Professional Programmer,

volume 25, pages 120–123. Miller Freeman Inc.

Chan, R. W. and Goldsmith, P. B. (2000). A psychovisually-

based image quality evaluator for jpeg images. In

Smc 2000 conference proceedings. 2000 ieee inter-

national conference on systems, man and cybernet-

ics.’cybernetics evolving to systems, humans, orga-

nizations, and their complex interactions’(cat. no. 0,

volume 2, pages 1541–1546. IEEE.

COMAU (2021). e.do robot. url: https://edo.cloud/

en/learning-center/. Accessed: 2022-07-20.

Court, R. E. C. (2022a). Como ocorre o teste de integri-

dade das urnas eletr

onicas. url= https://www.tre-

ce.jus.br/eleicao/eleicoes-2022/auditoria-de-funciona

mento-das-urnas-eletronicas/como-ocorre-a-auditoria

-de-funcionamento-das-urnas-eletronicas. Accessed:

2022-11-16.

Court, S. E. (2020). Processo eleitoral no brasil.

url= http://www.tse.jus.br/eleicoes/processo-eleitoral-

brasileiro. Accessed: 2022-10-10.

Court, S. E. (2022b). Ballot box. url= https://interna

tional.tse.jus.br/en/electronic-ballot-box/presentation.

Accessed: 2022-07-22.

Court, S. E. (2022c). Urna eletr

onica 25 anos: Lanc¸ado

em 1996, equipamento

E o protagonista da maior

eleic¸

ao informatizada do mundo. url= https://www.

tse.jus.br/comunicacao/noticias/2021/Maio/urna-ele

tronica-25-anos-lancado-em-1996-equipamento-e-

o-protagonista-da-maior-eleicao-informatizada-do-

mundo. Accessed: 2022-08-12.

Courtney, J. (2021). Sediqa: Sound emitting document im-

age quality assessment in a reading aid for the visually

impaired. In Journal of Imaging, volume 7, page 168.

MDPI.

Fachin, E. (2021). Resoluc¸

ao nº 23.673, de 14 de

dezembro de 2021. url= https://www.tse.jus.br/

legislacao/compilada/res/2021/resolucao-no-23-673-

14-de-dezembro-de-2021. Accessed: 2022-07-22.

Kastelan, I., Kukolj, S., Pekovic, V., Marinkovic, V., and

Marceta, Z. (2012). Extraction of text on tv screen us-

ing optical character recognition. In 2012 IEEE 10th

Jubilee International Symposium on Intelligent Sys-

tems and Informatics, pages 153–156. IEEE.

Kwon, K.-S. and Ready, S. (2014). Practical guide to ma-

chine vision software: an introduction with labview.

John Wiley & Sons.

Mariappan, M., Ramu, V., Ganesan, T., Khoo, B., and Vel-

lian, K. (2011). Virtual medical instrument for otorob

based on labview for acquiring multiple medical in-

strument lcd reading using optical character recogni-

tion. In 2011 International Conference on Biomed-

ical Engineering and Technology, IPCBEE vol, vol-

ume 11.

Mittal, A., Moorthy, A. K., and Bovik, A. C. (2012a). No-

reference image quality assessment in the spatial do-

main. In IEEE Transactions on Image Processing,

volume 21, pages 4695–4708. IEEE.

Mittal, A., Soundararajan, R., and Bovik, A. C. (2012b).

Making a “completely blind” image quality analyzer.

In IEEE Signal Processing Letters, volume 20, pages

209–212. IEEE.

Pech-Pacheco, J. L., Crist

obal, G., Chamorro-Martinez, J.,

and Fern

andez-Valdivia, J. (2000). Diatom autofocus-

ing in brightﬁeld microscopy: a comparative study. In

Proceedings 15th International Conference on Pattern

Recognition. ICPR-2000, volume 3, pages 314–317.

IEEE.

Qin, H. and El-Yacoubi, M. A. (2017). Deep representation

for ﬁnger-vein image-quality assessment. In IEEE

Transactions on Circuits and Systems for Video Tech-

nology, volume 28, pages 1677–1693. IEEE.

Shrivakshan, G. and Chandrasekar, C. (2012). A compari-

son of various edge detection techniques used in im-

age processing. In International Journal of Computer

Science Issues (IJCSI), volume 9. Citeseer.

Smith, R. (2007). An overview of the tesseract ocr engine.

In Ninth international conference on document anal-

ysis and recognition (ICDAR 2007), volume 2, pages

629–633. IEEE.

Tekin, E., Coughlan, J. M., and Shen, H. (2011). Real-time

detection and reading of led/lcd displays for visually

impaired persons. In 2011 IEEE Workshop on Appli-

cations of Computer Vision (WACV), pages 491–496.

IEEE.

Image Quality Assessment in the Context of the Brazilian Electoral System

451