GEN101: Introductory Artificial Intelligence Abu Dhabi University

E-Waste Revolution: Machine Learning for Better Recycling

Ahmed Khaled Sakr (1093123), Muhammad Aasif (1093004)

Abstract
The rapid increase of electronic devices has made e-waste management a critical
environmental challenge. In this project, we aimed to improve the efficiency of e-waste
categorization and recycling through machine learning techniques. We used binary
classification and multi-classification models for classification into various categories like
laptops, mobile phones and televisions.
According to our results, both approaches demonstrated promising accuracy rates;
however, the multi-classification model appeared to outperform the binary classification
models, attaining 100% accuracy with no false negatives. The binary classification models
though indicated class performance at an individual level in detail as well as near-perfect
classification by way of ROC and PRC curves. These findings show how machine learning
can change e-waste management and why it is important to select appropriate classification
methodologies that meet specific needs or objectives. In future studies, certain aspects may
be devoted to bettering classifiers while incorporating them into real-world recycling
systems to minimize the impact towards the environment due to the accumulation of E-
waste.

1. INTRODUCTION
Electronic waste is one of the rapidly growing waste sectors in the world. It has become a growing
environmental concern due to recent technological advancements. With 62 million metric tons
generated across the world in 2022, the volume of discarded electronic material has nearly doubled
since 2010 [1]. UAE is one of the highest e-waste-generating countries in the world. According to the
United Nations Global Waste Monitor Report from 2016, the per capita e-waste production in the UAE
amounted to 17.2 kg, with the total quantity of e-waste equaling 134,000 tons [3]. And these numbers
have been growing exponentially since then.
Managing and Recycling of E-waste remains inadequate, with over three-quarters of the e-waste still
going undocumented. This E-waste holds large volumes of raw materials including gold, palladium,
silver, iron, copper, and many rare-earth metals. However, e-waste recycling centers face challenges in
sorting and categorizing various types of electronic waste for appropriate recycling processes.
Our project aims to further improve the current E-waste management system by the implementation of a
machine learning model made using teachable Machines to differentiate different types of e-waste such
as Laptop, Mobile and Television. The Motivation behind this project is to improve the recycling
quality, accuracy and efficiency of the recycling centers. This project’s benefits are numerous. By
automating the categorization of the waste materials, we can enhance the recycling process, reduce the
workload on manpower and ensure that each device has undergone the right recycling process. This will
also help in the extraction of the valuable materials from these wastes contributing to a more sustainable
recycling process.
Overall, in our project, we demonstrate that the methods of AI may be successfully applied to confront
the issues that relate to the modern problem of e-waste and its recycling thus making the subsequent
technological advances more sustainable.

2. APPROACH
There were six phases conducted to complete this project. Figure 1 summarizes the overall process of
the project. The first step was to collect the raw data. We collected 70 images for each of the classes. In
the next step, we augmented the data by mirroring it to obtain more data from the collected data. Then,
we resized the images to 512x512 pixels. Then we prepared the data for training the model. We used the
test set cross-validation method taking 70% for the training set and 30% for the testing set. Afterwards,
we trained the models on the Teachable Machine using the training set. Finally, we evaluated the model
by testing each sample on the AI Model and getting the results. We constructed Precision-Recall Curve
and Receiver Over Characteristic Curve to evaluate the model.
Abu Dhabi University

Fig. 1: Phases of the AI Project.

2.1 Data Collection
Ahmed Khaled:

I collected 3 types of data in the form of an image (.jpeg format). I collected images of Laptops, Mobile
Phones, and TV’s. In total, I collected 105 Images. Then I put each of them into 3 different folders with
each different labelled class and all 3 contain 35 images of the specified class. And I collected data in 2
ways. I went to a local mall, and I asked for permission to take photos of the devices that were displayed
and made sure I had the green flag to do so, I started collecting images of the devices from the top and
the sides. I also did it while the device screen was on and off and for the laptops, I had the screen lid
closed and wide open and took images of it. And I also took some pictures of some of the devices I
possess. I did all these steps so we could make the perfect model with almost no mistakes. It was kind of
hard at first to persuade the mall security to allow us to take the images but when I told them the reason,
they allowed us to do so.

Muhammad Aasif
We were tasked with collecting image samples of Laptops, mobile phones, and Televisions to build our
model. My data collection method was mostly conducted with the help of my friends and family. I
visited some of my friends and family to take the images specifically with their permission. For the
television samples, I visited a Digital Store near me as I had a limited number of televisions using my
contacts. I asked the permission of the store to take the pictures, which was quite challenging, but with
an explanation of how the images would be used, they helped us. In total, I obtained 105 images (35
images per class) in jpg format.
To ensure the quality of the data, the pictures were taken with adequate lighting and using the best
camera at my disposal. I also made sure that there was no blur or noise in the images taken. I took the
best images from the ones I have taken so that the quality of the model produced is the best. The photos
were taken from different angles to give the machine model much more information. No more than 3
pictures were taken of the same object so that the model gets a different variety of data. The images of
the Laptop were taken with the lid open, closed and at an angle whereas the images of the Mobile phone
were taken at different angles of both the back side and the side. The images were imported to my
laptop with HD quality. I classified the images into folders for each of the respective classes. Data
Cleaning processes were applied to the data to ensure there were no duplicates, outliers or irrelevant
images among the data. It was quite challenging and fun to collect the data as we visited many places to
get it. I also made sure that the images were different and unique from the ones my teammates had
taken.
GEN101: Introductory Artificial Intelligence Abu Dhabi University

Data Augmentation and Resizing

Once we were done with the data collection phase, we combined all the samples. We used zip folders to
transfer the data as it will ensure that all the data will be transferred with much safety. We organized them
in folders such that we would not lose any data and it would be easier to access and work with it. We
classified them into their classes with utmost care to not let data be duplicated or missed. We then used the
Mirroring Augmentation technique on the data using a website named "Pinetools” to double the size of the
samples. We combined both the original and the augmented samples.
Later, we resized the images using the Bulk resize website to resize all the images to 512x512 pixel size as
the teachable machine model only works on square images and as this size gives better results at a faster
rate. At last, we randomized the samples obtained by randomly renaming them using a website named
“Web Rename”. Then, we renamed the data giving it a practical name in the same new order we got after
renaming using the website. This way we ensured the data was randomly selected for the later stages of
building the machine learning model.

We came to know that it is always better to take data from different places with different backgrounds and
different varieties of data. We tried to collect samples of different types of gadgets in the market at different
times. We've obtained pictures of gadgets from around 2004 to the current date. This will provide the
machine learning model with a variety of data to work with and produce the best-fit model. We came to
know that raw data collection is not an easy task as it requires a lot of time and effort, ort, especially when
we have a list of requirements to fulfil for each sample. From this project, we learned what type of data
collection and datasets would help us train a best-fit model. We comprehended how we can get more data
with whatever data we have using augmentation methods. We would have tried to get much better, diverse
and larger data to obtain a better version of the model.

Data Distribution - Multi Classifier

120

100
98 98 98
80

60

40
42 42 42
20

0
Laptop Mobile TV

Training Testing

Fig. 2: The Figure above shows how the Training and Testing data were distributed for the
Multi-classification model.
Abu Dhabi University

Data Distribution - Laptop Binary Classifier

120

100
98 98
80

60

40
42 42
20

0
Only Laptop Images Mobile & TV Images
Laptop Non - Laptop

Training Testing

Fig. 3: The Figure above shows how the Training and Testing data were distributed for the
Laptop - binary classification model.

Data Distribution - Mobile Binary Classifier

120

100
98 98
80

60

40
42 42
20

0
Only Mobile Images Laptop & TV Images
Mobile Non - Mobile

Training Testing

Fig. 4: The Figure above shows how the Training and Testing data were distributed for the
Mobile - binary classification model.
GEN101: Introductory Artificial Intelligence Abu Dhabi University

Data Distribution - TV Binary Classifier

120

100
98 98
80

60

40
42 42
20

0
Only TV Images Laptop &Mobile Images
TV Non - TV

Training Testing

Fig. 5: The Figure above shows how the Training and Testing data were distributed for the
TV - binary classification model.

Fig. 6: A Data Montage of our samples.

Abu Dhabi University

2.2 AI Model
Teachable Machine uses MobileNet as its Convolutional Neural Network (CNN)
MobileNet is a Convolutional Neural Network (CNN) that was open-sourced by Google, and it gives
an excellent starting point for training classifiers that are insanely small and fast. Mobile net reduces
the complexity of the computations using depth-wise separable convolutions. It splits the convolution
into 2 layers: ( i ) It uses a Depth-wise convolution to significantly decrease the number of parameters
which results in lightweight deep neural networks by applying a single filter to each of the input
channels, and ( ii ) a Pointwise convolution, which applies a 1x1 convolution to combine these
outputs. The operation of the depth-wise convolutions

Fig. 7: MobileNet Model behind Teachable Machine.

3. RESULTS, ANALYSIS AND INTERPRETATION OF DATA

We decided to use the Test-set method as the cross-validation method with 70% of the obtained data
for training the model and the rest 30% for testing the model.
Firstly, we made three binary classification models for each of the classes Laptops, Mobiles and
Television respectively. We used 280 images for each classification model taking 140 samples for the
positive class and 140 for the negative classes (equally split between the other two classes). The Data
distribution of Training and Testing of each classifier is shown in Figure 8.

Fig. 8: Data Distribution Tables for the binary classification models.

After splitting the data into train and test for each of the models, we proceeded to train the teachable
GEN101: Introductory Artificial Intelligence Abu Dhabi University
machine using the training data. Once the training of the model was completed, we started testing the
model using the test data one by one while recording the percentage of the positive class given by the
AI Model. We tested the model with all 84 test samples and recorded the results from the AI.
Once the testing phase was done we evaluated the model using the performance criteria: Accuracy,
Precision, Recall, Specificity, False Positive Rate, and Visual Results for ROC and PRC Curves using the
equation shown in Figure 10.
We varied the threshold percentage for the AI Model from each of the percentages to obtain the values of
the performance criteria at different threshold values. We used the following formula for obtaining the
prediction of the AI Model in Excel:
=IF(B3>=$B$3,"Laptop","Non - Laptop")
We changed the values of the threshold by changing the highlighted cell value from the first one to the last
one.

Fig. 9: Overview of the Binary Classification Model Excel sheet.

Fig. 10: Formulas for finding the Evaluation Metrics.

Abu Dhabi University

Fig. 11: Precision-Recall Curve for the Laptop Binary Classifier.

Fig. 12: Precision-Recall Curve for the Mobile Binary Classifier.

GEN101: Introductory Artificial Intelligence Abu Dhabi University

Fig. 13: Precision-Recall Curve for the TV Binary Classifier.

Fig. 14: Receiver Over Characteristics Curve for the Laptop Binary Classifier.
Abu Dhabi University

Fig. 15: Receiver Over Characteristics Curve for the Mobile Binary Classifier.

Fig. 16: Receiver Over Characteristics Curve for the TV Binary Classifier.

After wrapping up the Binary Classification models, we moved on to the Multi-classification model.
Even for this model we used the Test-set method for the cross-validation method. We used a total of
420 samples with 70% (98 samples per class) for training and the rest 30% (42 samples per class) for
testing the model. A detailed distribution of the data is shown in the figure. We calculated the
Accuracy from the constructed confusion matrix using the given formula. We divided the sum of the
yellow highlighted cells with the sum of all the cells of the confusion matrix. We got an accuracy of 1
and the training time it took for the multi-classifier to learn our data and train was 11 seconds.
GEN101: Introductory Artificial Intelligence Abu Dhabi University

Fig. 17: Data Distribution of the Multiclassification Model.

Fig. 18: Confusion Matrix obtained from the Multi-Classification Model.

By Evaluating the performance of the binary and the multi-classifier using tables in Excel, we
obtained some valuable insights into how well it classifies different materials. Using the binary
classification table, we obtained the percentages of the positive class for each of the binary classification
models. We varied the threshold percentage for prediction by the AI model and recorded the values of True
Positives, False Positives, False Negatives and True Negatives for each of the threshold percentages. We
calculated the performance evaluation for Accuracy, Precision, Recall, Specificity, False Positive Rate, and
Visual Results for ROC and PRC Curves.
On the other hand, the multi-classifier table allows us to get a better picture of how the model works with
all three classes together. In this, we only took the predicted class (percentage of 50 and above) from the
teachable machine as the output. We were able to understand how the model worked with different test data
with all the possible combinations.
The cross-validation technique used for this project was the test-set method. We chose this method as this is
one of the easiest and cheapest validation techniques to use. Also, as we had a good number of data samples
this cross-validation technique was suitable for evaluating the model. For this, we classified the data into 2
parts: 70% for training and 30% for testing. These samples were divided randomly ensuring that there were
no duplicates and missing samples.

Multi-classification Model Versus Binary Classification Models

Using the Binary classification Table with the Multiclassification Table we were able to learn the
differences between the two. In the binary classification model, we were able to obtain many parameters to
evaluate the classification model such as Accuracy, Precision, Recall, Specificity, False Positive Rate, and
Visual Results for ROC and PRC Curves. We were able to find these measures for different thresholds to
find the best threshold for distinguishing between the classes. On the other hand, In the multiclassification
model, we only yielded the Confusion Matrix, the Accuracy and the Model Complexity represented by the
training time of the model. The only common comparable measure for both types of models is Accuracy.
We obtained an accuracy of 100 per cent in the multiclassification model with no false classification by the
model. While in the binary classification, it varied from 94.04% (Laptop –Binary Classifier) to 98.81%
(Mobile – Binary Classifier and TV – Binary Classifier). This comparison of the accuracies shows that the
multi-classification model performed well when compared with the binary classification models. The binary
classification model for Laptops struggled a bit with classifying the laptops as the model confused the
samples with the TV because of the relative size and Mobile phones due to the angle of the pictures taken.
Additionally, the Receiver over Charesterics curve plotted for each of the binary classification models
yielded great results with the Area under the curve (AUC) almost equal to one. This suggests the model
distinguished between the positive and negative classes, resulting in a near-best-fit model. Also, the
Precision-Recall Curve (PRC) plotted for the binary classifiers shows the models are near best-fit models.
Abu Dhabi University
The multi-classifier table allowed us to get a bigger picture of how the model distinguished and
classified the samples into each of their respective classes. The model showed excellent accuracy in
distinguishing the classes with no false classifications. This infers that our multi-classifier worked well
rather than the binary classification model.

CONCLUSION
In conclusion, our project aims to handle the major concern over e-waste management and recycling,
especially in the UAE with Machine Learning techniques. To improve the recycling process, we opted
for binary classification and multi-classification models.
Through many tests and evaluations, there were significant differences between binary classification
and multi-classification models. Both models however showed promising results yet the accuracy of
the multiclassification model was better than that of binary classification models. Despite these
challenges, binary classification models had a high accuracy rate ranging from 94.04% up to 98.81%.
Individual classes such as laptops, mobile phones, and televisions can be well explained by using
binary classification models. However, certain categorizations like laptops were tricky due to size
angle among other things.
Conversely, the multi-classification model gave an elaborate view of its classification through which it
achieved perfect accuracy without any false classifications. It could therefore effectively differentiate
between all three types of electronic devices including Laptops, TVs, and Mobile Phones.

REFERENCES
[1] Topic: Electronic waste worldwide. (2024, May 6). Statista.
https://www.statista.com/topics/3409/electronic-waste-worldwide/#topicOverview
[2] Global e-waste generation 2010-2022 | Statista. (2024, April 25). Statista.
https://www.statista.com/statistics/499891/projection-ewaste-generation-worldwide/
[3] Waheed, K. A., Singh, A., Siddiqui, A., Gamal, M. E., & Laeequddin, M. (2023). E-
Waste3333333333,,,,3, recycling behaviour in the United Arab Emirates: Investigating the
roles of environmental consciousness, cost, and infrastructure support. Sustainability, 15(19),
14365. https://doi.org/10.3390/su151914365
[4] NN SVG. (n.d.). https://alexlenail.me/NN-SVG/AlexNet.html
[5] Iqbal, H., & Iqbal, H. (2024, April 8). E-waste Recyclers are a Growing Concern to the
Economy of the UAE. EcoMENA. https://www.ecomena.org/e-waste-recyclers-a-growing-
concern-to-economy-of
uae/#:~:text=The%20government%20of%20UAE%20has,on%20behalf%20of%20the%20gov
ernment.