Application of Computer Vision Technologies for Automated Utility

Meters Reading

Maria Spichkova and Johan Van Zyl

School of Science, RMIT University, Melbourne, Victoria 3000, Australia

Keywords:

Software Engineering, Computer Vision Techniques, Case Study.

Abstract:

This paper presents a study on automated reading of utility meters using two computer vision techniques:

an open-source solution Tensorﬂow Object Detection (Tensorﬂow) and a commercial solution Anyline. We

aimed to identify the limitations and beneﬁts of each solution applied to utility meters reading, especially

focusing on aspects such as accuracy and inference time. Our goal was to determine the solution that is the

most suitable for this particular application area, where there are several speciﬁc challenges.

1 INTRODUCTION

Development and application of smart meters is an ac-

tive research topic over the last decades, see (Depuru

et al., 2011; Benzi et al., 2011; Zheng et al., 2013).

The smart devices are elaborated for several types of

utilities, e.g., electricity, gas, etc. Their advantage is

that the data on energy consumption is recorded and

sent automatically to the provider as well as, in some

cases, to the corresponding customer. This data can

be used for monitoring and billing purposes. This al-

lows more detailed analysis of the consumption pat-

terns as well as the ways to reduce consumption or

to schedule the energy-consuming tasks for the time,

which is mostly suitable for the energy network (in

the terms of payment or the energy load).

The core disadvantage of this solution is that its

implementation on a large scale is expensive. In the

case customers have to pay for an upgrade to a smart

meter, they might prefer to object the upgrade. Some

customers perceive that the learning curve for using

smart meters is steep, and prefer to avoid using them

by this reason. Also, some customers are concerned

regarding their privacy while using the smart meters,

as the information regarding the usage pattern over

the day might be used to identify whether the resi-

dents are currently at home, home many of them are

at home, etc. Therefore, in some countries the roll out

to the smart meter systems is done step-wise, where

on the initial stage the roll out is voluntary.

On the other hand, a manual collection of meter

readings is not only time consuming, but also some-

times complicated for people with vision impairment.

In our earlier work (Spichkova et al., 2019b), we con-

ducted a project in collaboration with Energy Aus-

tralia, which is an electricity and gas retailing private

company that supplies electricity and natural gas to

more than 2.6 million residential and business cus-

tomers throughout Australia. Their solution for non-

smart meters was to provide an online portal, where

the consumers can update the records on the utility

readings. Thus, the consumers had to provide a lot of

additional details, and to calculate their utility read-

ings manually. The goal of our previous project was

to elaborate an alternative method for the existing sys-

tem, which would allow for a higher degree of au-

tomation to increase the usability of the system. The

proposed solution was to use computer vision tech-

niques for capturing readings. We analysed there

the following computer-vision technologies: Google

Cloud Vision

(GCV), Amazon Web Services (AWS)

Rekognition

, Tesseract OCR (Smith, 2007), and

Azure’s Computer Vision

. The study demonstrated

that AWS Rekognition provides better results for this

application domain. However, it’s accuracy was far

from ideal: the average accuracy values AWS Rekog-

nition was only 36%.

Contributions: In this paper we present our recent

study, where we compared with our previous results

two further computer vision technologies, Tensorﬂow

Object Detection (Tensorﬂow) and Anyline:

https://cloud.google.com/vision

https://aws.amazon.com/rekognition

https://docs.microsoft.com/en-us/azure/

cognitive-services/computer-vision

Spichkova, M. and Van Zyl, J.

Application of Computer Vision Technologies for Automated Utility Meters Reading.

DOI: 10.5220/0009892505210528

In Proceedings of the 15th International Conference on Software Technologies (ICSOFT 2020), pages 521-528

ISBN: 978-989-758-443-5

521

• Tensorﬂow is an open-source machine learning

system that operates at large scale and offers

a multitude of models to be retrained (more

than 30), see (Abadi et al., 2016; Abadi et al.,

2017). To analyse the Tenserﬂow results, we ap-

plied the corresponding visualisation tool Tensor-

Board (Wongsuphasawat et al., 2018) that allows

to visualise TensorFlow graphs, plot correspond-

ing quantitative metrics, etc.

• Anyline is a commercial solution intended to read

utility meters, which also offers a free sample app

Anyline OCR Scanner

that has been used during

tests within our project.

The obtained results are signiﬁcantly better (see Sec-

tion 4 for details) in the terms of recognition accuracy

than the results of our early investigation study within

the domain of meter reading recognition (Spichkova

et al., 2019b), where we analysed another two com-

puter vision technologies, applied to the same data

sets: the results of the current study demonstrate the

accuracy up to 92.35%.

(a) Original image with reﬂections.

(b) Application of the thresholding technique.

Figure 1: Reﬂection issues.

2 BACKGROUND

There are many challenges for application of com-

puter vision technologies for the reading of utility me-

ters, such as:

• Reﬂections: Most meter models encountered has

a transparent protective cover, see Figure 1(a),

which lead to reﬂections from it. This be-

comes problematic for computer vision technolo-

gies which includes thresholding/ﬂooding tech-

niques.

https://www.tensorﬂow.org/guide/summaries and

tensorboard

https://anyline.com/products/ocr-meter-reading

Thresholding techniques are usually applied to

OCR technologies in-order to minimize noise and

or to convert to black-and-white images. Fig-

ure 1(b) illustrates an application of threshold.

The threshold in this case has removed the last

digit completely and clipped several other digits.

Flooding is a technique used to ﬁnd similar neigh-

bouring pixels. This technique can be used when

ﬁnding contours of shapes. Figure 1(c) displays

ﬂooding made on the last digit. This technique

failed to ﬁnd the contour of the digit and the con-

tour of reﬂection was found instead.

• Clipped Digits: The ﬁnal digit in analogue meters

usually rotate freely. This becomes problematic as

the digit becomes clipped and the full digit is not

displayed. Figure 2 presents an example of such

case: the ﬁnal digit is both a 2 and a 3 with neither

being displayed completely. The computer vision

technology would need to be able to read clipped

digits.

Figure 2: Clipped Digits.

• Not all characters and digits, which can be iden-

tiﬁed on the meter, actually belong to the meter

reading: Utility meters commonly include other

text as seen red boxes in Figure 3. The computer

vision technology would need to be able to dis-

criminate which digits are part of the meter read-

ing, and which have to be ignored.

Figure 3: Digits on the meter, which do not belong to the

meter reading.

• Blur, noise, warping, etc.: The grime adds noise

to the observed meter. Grime with reﬂection cre-

ates a blur effect around some digits, see Figure 4

(digits 2 and 5). Digits are also observed to be

warped along with the shape of the cover, e.g., the

digits appear “stretched” or “squashed” depend-

ing on observation angle. Furthermore, there can

be a signiﬁcant contrast difference in scenarios

where the colour of the digits is mixed. For exam-

ple, in Figure 4, the ﬁnal 3 digits (883) have low

ICSOFT 2020 - 15th International Conference on Software Technologies

522

contrast and are barely observable in comparison

to the ﬁrst 4 digits.

• Different representation styles (scales, dials and

digits) mixed within a meter interface.

Utility models observed during the project pre-

sented their reading value either through: using

rotational dials, using cyclometers, a combination

of rotational dials and cyclometers or a singular

digital display. The challenge becomes even more

complex when the numeric value is displayed on

a scale, e.g., as presented in Figure 5:

– The meter (a) should be read as 5762.615m

gas and not 57626,

– The meter (b) should be read as 14485.68kWh

energy and not 14485,

– The meter (c) should be read as 75691.1kWh

energy and not 756911.

The meters commonly include a decimal point.

Most meters include the decimal point as the last

digit and can be read and a tenth. Sometimes it is

a mixture of a digit and a dial as in Figures 5(b)

and (c).

Some models display three decimal points as seen

in Figure 4 (the digits after decimal points are

highlighted with red colour), which should read

as 2550.883m

and not 2550883. Thus, our aim

is to identify a computer vision technology that

is capable of distinguishing between different nu-

meric scales and be able to detect both digits and

rotation of dials.

3 METHODOLOGY

To determine which of these technologies is most suit-

able for reading utility meters, we elaborated a set of

tests that allows us to identify the limitations of each

technology by gradually adjusting image blur, noise,

gamma or scale.

The following methodology was applied to anal-

yse the techniques:

1. To create a training dataset for the Tensorﬂow Ob-

ject Detection framework.

2. To train all the different models using the dataset,

which was elaborated at Step 1. This was done

through Google ML Engine.

3. To create evaluation datasets for which the tech-

nologies can be tested against.

4. To create a test harness for the involved technolo-

gies.

5. To run the test harness on the evaluation datasets

created from Step 3.

The training dataset is a set of all images found during

project duration. The ﬁnal training dataset consisted

of 395 images and 2000 annotations. Unfortunately,

this is still considered limited as supplied Tensorﬂow

models were created based upon 2000+ annotations

per object.

As the evaluation dataset we used the same im-

ages as in our earlier work (Spichkova et al., 2019b),

which allowed us to compare the results of applica-

tion of Tensorﬂow and Anyline not only with each

other, but also with the results of AWS Rekognition

(which demonstrated the best but not good enough

accuracy in our previous study). Thus, in (Spichkova

et al., 2019b), a total of 30 images were selected based

on their “uniqueness” – images with unique meters

or images with unique lighting. These images were

duplicated and modiﬁed with various effects in or-

der to test the limitations of the different technologies.

These effects are:

• Scaling: The dataset was scaled in steps of 0.1

ranging from a scale of 0.1 to 0.9 (10% to 90%)

of the original dataset.

• Blurring: Blurring was done in steps of 10 from

10 to 90 with an open source blur algorithm that is

based on the normalised box ﬁlter, see (OpenCV,

2018). The algorithm uses a normalised box ﬁlter,

the numeral value adjusts the kernel size.

• Gamma: The gamma algorithm was used with

an open source lookup table algorithm (OpenCV,

2018). The gamma correction to simulate differ-

ent lightning conditions.

• Noise: The noise algorithm is based upon the

salt and pepper noise algorithm that adds sharp

and sudden disturbances in the image in the form

of sparsely occurring white and black pixels, see

(Gonzalez and Woods, 2001). This algorithm was

included to further test the performance of the

various technologies as noise arguably emulates

“dirt” on meters.

In contrast to Anyline, Tensorﬂow’s Tensorboard

provides more in-depth evaluation of each model and

how well each model detects objects for a given

dataset, such as:

• single-shot multi-box detector (SSD), see (Liu

et al., 2016),

• feature pyramid networks (FPN), see (Lin et al.,

2017),

• fast region-based convolutional neural networks

(FRCNN), see (Ren et al., 2017).

Application of Computer Vision Technologies for Automated Utility Meters Reading

523

(a) (b)

Figure 4: Challenging case: (a) Original image; (b) Application of the thresholding technique over the original image.

(a) (b) (c)

Figure 5: Different representation styles (scales, dials and digit) mixed within a meter interface.

Our initial hypotheses in terms of accuracy and detec-

tion rates were as follows:

H1: FRCNN model would signiﬁcantly outper-

form other models.

H2: SSD model would perform signiﬁcantly

worse than the other models in terms of accuracy.

H3: The lower the image resolution is, the faster

inference would occur.

H4: FRCNN would require signiﬁcantly more

time than FPN and SSD models.

As Tensorﬂow Object Detection only detects ob-

jects, the results have to be ﬁltered in order to give a

reading. We applied the ﬁltering algorithm presented

in Algorithm 1 to

• remove any junk data, i.e., any detected objects

for what the identiﬁcation conﬁdence is ≤ 10%,

• remove all duplicates within the geometric region,

keeping the results withe the highest identiﬁcation

conﬁdence.

The ﬁltering algorithm can be further improved, if the

previous reading of the meter is considered, for which

an access to the customer’s account is required. We

haven’t applied this improvement within the compar-

ison study, as the supplied Anyline app cannot include

any customer related data.

Results from each dataset and each technology

were produced in csv ﬁles (a comma-separated values

ﬁle that allows data to be saved in a tabular format)

with the following structure:

FileName, InferTime, FilteredReading,

ExpectedReading, IsCorrect

where InferTime denotes the interference time,

which was measured in ms.

Similarly to (Spichkova et al., 2019b), we calcu-

lated the accuracy of recognition calculated as the fol-

lowing simple formula (we measure the accuracy in

percents, where 100% means a totally accurate recog-

nition):

Accuracy =

CorrectResults

Total

∗ 100 (1)

where:

CorrectResults is the number of results that match

with the original readings completely,

Total presents the total number of images in a dataset.

As in our study, we had 30 images in each of the

datasets, Total = 30.

4 RESULTS AND DISCUSSION

Figures 7-9 present the identiﬁed accuracy scores per

dataset, where 50% indicates half of the dataset meter

images were correctly read.

It is important to mention that both Tensorﬂow

and Anyline were less sensitive to scaling, gamma,

and sat and paper issues than to blurring, but, the ac-

curacy of Tensorﬂow was almost twice higher:

• With low blurring (10BLUR and 20BLUR), Ten-

sorﬂow performed with 100% accuracy, where the

accuracy of Anyline dropped to approx. 63% and

50% respectively.

• For the effect of 50BLUR, the accuracy of Ten-

sorﬂow FPN and Anyline were approx. 87% and

37% respectively.

ICSOFT 2020 - 15th International Conference on Software Technologies

524

Algorithm 1: Filtering of meter readings.

1: j = 1

2: for all i ← 1, n do

3: if con f idence(result[i]) > 10% then

4: if result[i] is unique within geometric region then

5: list[ j] = result[i]

6: j = j + 1

7: else  where list[k] is the duplicate of result[i]

8: if conﬁdence(result[i]) > conﬁdence(list[k]) then

9: list[k] = result[i]

10: end if

11: end if

12: end if

13: end for

14: Sort list[i] by geometric location from left to right

Figure 6: Example of an application of the blurring effect

with 90BLUR.

• For the effect of 90BLUR, the accuracy for Ten-

sorﬂow FRCNN, Tensorﬂow SSD and Anyline

was only 10%, where the accuracy of Tensorﬂow

FPN was approx. 33%. However, the 90BLUR

effect means a very blurry image, see Figure 6 for

an example.

The overall performance of the Tensorﬂow models

greatly surpass expectations in terms of accuracy,

having an average performance of 88.14%, 89.51%

and 92.35% for FRCNN, SSD, and FPN, respectively.

This is especially remarkable, if we compare it with

the average accuracy values from AWS Rekognition

was only 36% that demonstrated the best (but not re-

ally satisfactory) results within the study presented in

(Spichkova et al., 2019b).

TensorBoard conﬁrms the accuracy of the trained

models. Scoring a near perfect score of 1.0 is ex-

tremely signiﬁcant is a strong indication that Tensor-

ﬂow Object Detection is a suitable framework for the

automated meter reading.

Anyline performed arguably well having an aver-

age performance of 57.16%, and struggled on several

utility meter models. The results indicate that Anyline

may not have trained or tested their product on a simi-

lar utility meters models as used within the Australian

Figure 7: Accuracy scores per dataset: Gamma analysis.

market. The accuracy of Anyline was signiﬁcantly

lower for all data sets: if compared with Tensorﬂow

FPN, the accuracy of Anyline was in average lower

approx. 35% lower, where

• the largest differences in the cases of

0.25GAMMA (approx. 70%) and 20BLUR

(approx. 57%);

• the largest difference (approx. 23%) was in the

cases of 1.25GAMMA, 0.12SP (noise), 90BLUR,

and the original data sets.

With respect to our hypotheses H1-H4, the results of

the conducted study can be summarises as follows:

• H1 and H2 were disproved by the conducted

study: In the terms of accuracy, the best per-

forming model was FPN, where FRCNN and SSD

were performing slightly worse than the other

models.

• H3 was proved as correct.

• H4 was also proved as correct: In average, FR-

CNN was approx. 2.5 slower than FPM and ap-

prox. 3.2 times slower than SSD.

The results for both Tensorﬂow and Anyline are

also signiﬁcantly better than the results of our

Application of Computer Vision Technologies for Automated Utility Meters Reading

525

(a) (b)

Figure 8: Accuracy scores per dataset: (a) Salt and pepper analysis; (b) Blur analysis.

Figure 9: Accuracy scores per dataset: Scale analysis.

early investigation study conducted for the domain

of meter reading recognition, where we analysed

Google Cloud Vision (GCV) and Amazon Web Ser-

vices (AWS) Rekognition using the same data sets.

The average accuracy values for GCV and AWS

Rekognition were just 29% and 36% respectively,

see (Spichkova et al., 2019b).

5 RELATED WORK

The research on the smart meter devices and the

corresponding analytic was actively conducted over

many years, which was reﬂected not only in research

publications but also in patents, see e.g., (Ehrke et al.,

2003; Grady et al., 2016; Winter, 2017).

Over the last decade, there were two core research

directions in this area: (1) privacy and security

aspects of the smart meter application, and (2) smart

meters in combination with a smart grid system. In

the rest of the section we discuss the most cited (as

per Google Scholar) publications, grouped by the

research directions.

Privacy and Security Aspects of Smart Meters:

This research direction is currently the most active

one among the mentioned directions, because the pri-

vacy and security concerns provide one of the biggest

obstacles for the (potential) users of smart meters.

In many cases, data mining and data analytics tech-

niques were applied on the meter reading data, to in-

vestigate the above issues questions.

A privacy-preserving smart meter architecture

was presented in (Molina-Markham et al., 2010). The

authors also conducted a study to demonstrate that the

power consumption patterns can help to reveal how

many people are in the home, what are their sleeping

and eating routines, etc.

A theoretical framework to analyse privacy as-

pects of smart meters was introduced in (Sankar et al.,

2013). A formal framework to quantify the privacy

trade off problem in smart meter data was introduced

in (Rajagopalan et al., 2011).

An approaches for occupancy detection from

electricity consumption data were proposed in

(Kleiminger et al., 2013).

The extraction of the households characteristics

from the the smart meter data was discussed in

(Beckel et al., 2014).

Similarly, the question on what the consumption

patterns (created on the basis of the smart meter data)

might say about the consumers, was discussed in (Al-

bert and Rajagopal, 2013) and (Beckel et al., 2014).

An approach for non-intrusive occupancy mon-

itoring using smart meters was discussed in (Chen

et al., 2013). This work aimed to implement energy-

efﬁciency optimisations based on the information of

home’s occupancy.

Design of Smart Meters for the Smart Grid: An

approach for anonymizing the data sent by a smart

meter to achieve security and privacy of the smart grid

was proposed in (Efthymiou and Kalogridis, 2010).

A smart meter data aggregation approach for smart

grids was introduced in (Li et al., 2010). The authors

ICSOFT 2020 - 15th International Conference on Software Technologies

526

applied homomorphic encryption to solve the privacy

issue. An overview of typical smart meter’s aspects

and functions wrt. smart grid aspects was presented

in (Zheng et al., 2013).

6 CONCLUSIONS

We presented the results of a research project, which

goal was to provide an alternative method for the cur-

rent system to update the meter reading data, collected

from non-smart utility meters.

Our early investigation study on the recognition

accuracy of Google Cloud Vision and AWS Rekogni-

tion applied for recognition in utility meter readings,

demonstrated very low average accuracy values (29%

and 36%, respectively). For this reasons, we con-

ducted a further study to analyse two other computer

vision technologies, applied for recognition in utility

meter readings:

• an open-source Tensorﬂow technique (FRCNN,

FPN, and SSD models), and

• a commercial solution Anyline.

The study demonstrated that Tensorﬂow provides sig-

niﬁcantly better results for our application domain

(92.35% for the FPN model), in comparison to Any-

line, as well as to Google Cloud Vision and AWS

Rekognition.

This research project was conducted under

the initiative Research embedded in teaching, see

(Spichkova, 2019; Simic et al., 2016; Spichkova

and Simic, 2017). This initiative was proposed at

the RMIT University (Melbourne, Australia) within

the Software Engineering projects (SEPs) conducted

in collaboration with industrial partners. The aim

of this initiative is to encourage students’ curiosity

for Software Engineering and Computer Science re-

search. To reach this aim we include research com-

ponents as bonus tasks in the ﬁnal year projects (on

both undergraduate and postgraduate levels), which

typically focus on software and system development.

Few weeks long research projects have been spon-

sored by industrial partners, who collaborated with

the students and academic advisers through the ﬁ-

nal year projects. Respectively, the topics of these

short research projects focus align the topics ﬁnal year

projects. The successful results of this initiative are

presented in (Christianto et al., 2018; Clunne-Kiely

et al., 2017; Spichkova, 2018; Spichkova et al., 2018;

Spichkova et al., 2019b; Sun et al., 2018; Chugh et al.,

2019; Gaikwad et al., 2019; Spichkova et al., 2019a).

ACKNOWLEDGEMENTS

We would like to thank Shine Solutions and Energy

Australia for sponsoring this project under the re-

search grant RE-03615.

REFERENCES

Abadi, M., Barham, P., Chen, J., Chen, Z., Davis, A., Dean,

J., Devin, M., Ghemawat, S., Irving, G., Isard, M.,

et al. (2016). Tensorﬂow: a system for large-scale

machine learning. In OSDI, volume 16, pages 265–

283.

Abadi, M., Isard, M., and Murray, D. G. (2017). A compu-

tational model for tensorﬂow: an introduction. In Ma-

chine Learning and Programming Languages, pages

1–7. ACM.

Albert, A. and Rajagopal, R. (2013). Smart meter

driven segmentation: What your consumption says

about you. IEEE Transactions on power systems,

28(4):4019–4030.

Beckel, C., Sadamori, L., Staake, T., and Santini, S. (2014).

Revealing household characteristics from smart meter

data. Energy, 78:397–410.

Benzi, F., Anglani, N., Bassi, E., and Frosini, L.

(2011). Electricity smart meters interfacing the house-

holds. IEEE Transactions on Industrial Electronics,

58(10):4487–4494.

Chen, D., Barker, S., Subbaswamy, A., Irwin, D., and

Shenoy, P. (2013). Non-intrusive occupancy moni-

toring using smart meters. In Embedded Systems For

Energy-Efﬁcient Buildings, pages 1–8. ACM.

Christianto, A., Chen, P., Walawedura, O., Vuong, A.,

Feng, J., Wang, D., Spichkova, M., and Simic, M.

(2018). Enhancing the user experience with vertical

transportation solutions. Procedia Computer Science,

126:2075–2084.

Chugh, R., Chawla, N., Gracias, R. M., Padda, J. S., Li,

S., Nguyen, M. T., Spichkova, M., and Mantri, N.

(2019). Automated gathering and analysis of cannabi-

noids treatment data. Procedia Computer Science,

159:2570–2579.

Clunne-Kiely, L., Idicula, B., Payne, L., Ronggowarsito, E.,

Spichkova, M., Simic, M., and Schmidt, H. (2017).

Modelling and implementation of humanoid robot be-

haviour. In Knowledge-Based and Intelligent Infor-

mation & Engineering Systems, pages 2249–2258. El-

sevier Science Publishers BV.

Depuru, S., Wang, L., Devabhaktuni, V., and Gudi, N.

(2011). Smart meters for power grid. challenges, is-

sues, advantages and status. In 2011 IEEE/PES Power

Systems Conference and Exposition, pages 1–7. IEEE.

Efthymiou, C. and Kalogridis, G. (2010). Smart grid pri-

vacy via anonymization of smart metering data. In

International Conference on Smart Grid Communica-

tions, pages 238–243. IEEE.

Application of Computer Vision Technologies for Automated Utility Meters Reading

527

Ehrke, L. A., Nap, K. A., and Dresselhuys, D. R. (2003).

Electronic electric meter for networked meter reading.

US Patent 6,538,577.

Gaikwad, P. K., Jayakumar, C. T., Tilve, E., Bohra, N., Yu,

W., and Spichkova, M. (2019). Voice-activated solu-

tions for agile retrospective sessions. Procedia Com-

puter Science, 159:2414–2423.

Gonzalez, R. C. and Woods, R. E. (2001). Digital Im-

age Processing. Addison-Wesley Longman Publish-

ing Co., Inc., 2nd edition.

Grady, B. D., Vaswani, R., and Pace, J. (2016). Method and

system of reading utility meter data over a network.

US Patent 9,464,917.

Kleiminger, W., Beckel, C., Staake, T., and Santini, S.

(2013). Occupancy detection from electricity con-

sumption data. In ACM Workshop on Embedded Sys-

tems For Energy-Efﬁcient Buildings, pages 1–8. ACM.

Li, F., Luo, B., and Liu, P. (2010). Secure information ag-

gregation for smart grids using homomorphic encryp-

tion. In Smart Grid Communications, pages 327–332.

IEEE.

Lin, T.-Y., Doll

ar, P., Girshick, R., He, K., Hariharan, B.,

and Belongie, S. (2017). Feature pyramid networks

for object detection. In CVPR, volume 1, page 4.

Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S.,

Fu, C.-Y., and Berg, A. C. (2016). Ssd: Single shot

multibox detector. In European conference on com-

puter vision, pages 21–37. Springer.

Molina-Markham, A., Shenoy, P., Fu, K., Cecchet, E., and

Irwin, D. (2010). Private memoirs of a smart meter.

In ACM workshop on embedded sensing systems for

energy-efﬁciency in building, pages 61–66. ACM.

OpenCV (2018). Open source computer vision. https://

docs.opencv.org/3.1.0.

Rajagopalan, S. R., Sankar, L., Mohajer, S., and Poor, H. V.

(2011). Smart meter privacy: A utility-privacy frame-

work. In International conference on smart grid com-

munications, pages 190–195. IEEE.

Ren, S., He, K., Girshick, R., and Sun, J. (2017). Faster

r-cnn: towards real-time object detection with region

proposal networks. IEEE Transactions on Pattern

Analysis & Machine Intelligence, (6):1137–1149.

Sankar, L., Rajagopalan, S. R., and Mohajer, S. (2013).

Smart meter privacy: A theoretical framework. IEEE

Transactions on Smart Grid, 4(2):837–846.

Simic, M., Spichkova, M., Schmidt, H., and Peake, I.

(2016). Enhancing learning experience by collabo-

rative industrial projects. In ICEER 2016, pages 1–8.

Western Sydney University.

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

In Conference on Document Analysis and Recogni-

tion, volume 2, pages 629–633. IEEE.

Spichkova, M. (2018). Automated analysis of the impact

of weather conditions on medicine consumption. In

25th Australasian Software Engineering Conference

(ASWEC), pages 166–170. IEEE.

Spichkova, M. (2019). Industry-oriented project-based

learning of software engineering. In ICECCS, pages

51–61. IEEE.

Spichkova, M., Bartlett, J., Howard, R., Seddon, A., Zhao,

X., and Jiang, Y. (2018). SMI: Stack Management

Interface. In 23rd Int. Conference on Engineering of

Complex Computer Systems), pages 156–159.

Spichkova, M. and Simic, M. (2017). Autonomous systems

research embedded in teaching. In International Con-

ference on Intelligent Interactive Multimedia Systems

and Services, pages 268–277. Springer.

Spichkova, M., van Zyl, J., Sachdev, S., Bhardwaj, A., and

Desai, N. (2019a). Comparison of computer vision ap-

proaches in application to the electricity and gas meter

reading. In ENASE, pages 303–318. Springer.

Spichkova, M., van Zyl, J., Sachdev, S., Bhardwaj, A., and

Desai, N. (2019b). Easy mobile meter reading for

non-smart meters. In 14th Int. Conf. on Evaluation

of Novel Approaches to Software Engineering. IEEE.

Sun, C., Zhang, J., Liu, C., King, B. C. B., Zhang, Y., Galle,

M., Spichkova, M., and Simic, M. (2018). Software

development for autonomous and social robotics sys-

tems. In Intelligent Interactive Multimedia Systems

and Services, pages 151–160. Springer.

Winter, D. (2017). Methods and systems of reading utility

meters and methods and systems of transmitting util-

ity meter data. US Patent 9,752,895.

Wongsuphasawat, K., Smilkov, D., Wexler, J., Wilson, J.,

Man

e, D., Fritz, D., Krishnan, D., Vi

egas, F. B.,

and Wattenberg, M. (2018). Visualizing dataﬂow

graphs of deep learning models in tensorﬂow. IEEE

transactions on visualization and computer graphics,

24(1):1–12.

Zheng, J., Gao, D. W., and Lin, L. (2013). Smart meters

in smart grid: An overview. In Green Technologies

Conference, pages 57–64. IEEE.

ICSOFT 2020 - 15th International Conference on Software Technologies

528