A feature-fusion transfer learning method as a basis to
support automated smartphone recycling in a circular
Nermeen Abou Baker [0000-0002-9683-5920] 1, Paul Szabo-Müller [0000-0002-6626-9390] 1
and Uwe Handmann [00 00-0003-1230-9446] 1
1 Computer Science Institute, Ruhr West University of Applied Sciences,
Bottrop Lützowstrasse 5, 46236, Germany
Abstract. In this paper, we present how Artificial Intelligence (AI) could support
automated smartphone recycling, hence, act as an enabler for Circular Smart Cit-
ies (CSC), where the Smart City paradigm could be linked to the Circular Econ-
omy (CE), which is a leading concept of the sustainable economy. While business
and society strive to gain benefits from automation, the ongoing rapid digitaliza-
tion, in turn, accelerates the mass production of Waste Electric and Electronic
Equipment (WEEE), often called E-Waste. Therefore, E-Waste is the fastest
growing waste stream in the world and comes up with several negative environ-
mental and social impacts. In our research, we show an AI technique (particu-
larly, Transfer Learning) that could become an enabler for the CSC and the CE
in general and supporter of automated recycling, specifically. However, research
on this topic is emerging only recently, and practical applications are lacking
even more. For instance, object recognition has extensive research, whereas
smartphone classification nevertheless has rare attention. Our main contribution
is a Transfer Learning (TL) approach based on visual-feature extraction to clas-
sify smartphones; as a result, it supports automated smartphone recycling inde-
pendently of brands and even without any ex-ante information about product de-
signs. Our findings show that the main advantages of using TL, are reducing the
size of the training-set, computation time, and significant enhancements without
designing a completely new network from scratch. This may ease the automated
recycling of smartphones as well as other E-Waste, hence, contribute to the de-
velopment of the CE and CSC.
Keywords: Feature Fusion, Transfer Learning, Smartphone Recycling, Circular
Economy, Automation Systems, Smart City, Sustainability, E-Waste Manage-
ment, Circular City.
1.1 Motivation and Challenges
The interplay of emerging digital technologies such as AI, Smart City development, CE
opportunities, and challenges associated with E-Waste brings us to our research ques-
tion Fig. 1: How can AI (particularly, TL) be applied in order to enable automated
smartphone recycling, hence, contribute to the development of CSC?
In particular, this paper addresses the problem of smartphone recycling and applies
a feature-fusion TL method to classify smartphones without any ex-ante information
about product designs. In our interdisciplinary research in cooperation with digitaliza-
tion and sustainability, we embed this deep investigation in the wider framework of
Smart City development and CE.
Cities around the world are looking for strategies to become more sustainable places.
On one hand, economic prosperity, environmental quality, and social wellbeing should
go hand in hand. On the other hand, cities try to cope with global and local challenges,
such as; climate change, air pollution, biodiversity loss, social inequality, and resource
depletion. These visions of sustainable city convergence with digital technologies, like
AI, 3D Printing, Big Data Analysis, and the Internet of Things (IoT) in the smart city
concept and almost all areas of life [1–3].
Fig. 1. Our cooperative interdisciplinary research with digitalization and CE in the framework of
Particularly, AI could become the fundamental driver of CE and CSC. Despite that,
the smart city concept faces some challenges concerning the security and privacy is-
sues, and the rising of infrastructure costs, there are still ubiquitous areas of application,
such as; enhancing the city’s security level by recognizing people’ faces [44, 45] to
access restricted areas [8–10], improving traffic flows by partly autonomous drones and
vehicles [11–13], traffic management and smart tracking, assistance systems [14, 15],
predictive maintenance [16, 17], and last but not least, smart waste management, such
as ; installing sensors on waste bins to enhance the collection, smart disposal seg-
regation, sorting and disassembling, and maximizing materials use.
Some smart city initiatives also aim to become circular cities by picking up elements
of the CE, to magnify benefits from smarter use of resources . The CE concept pro-
poses low-emission and resource-saving modes of production and consumption by
closing material loops and extending product life-cycles. In the combination of the
smart city and the CE concept, we see a kind of new category or focus of action, re-
spectively, which we call a Circular Smart City (CSC).
In general, digital technologies could pull down some existing barriers to the CE,
like lacking knowledge about the location and condition of obsolete products or in-
cluded as well as currently higher costs of their treatment compared to ‘non-circular’
ones [5, 6]. By doing so, digitalization could support the application of CE strategies,
for example, some of the so-called R-Strategies like the redesign, reuse, redistribution,
refurbishment and maintenance, repair, remanufacturing, as well as recycling of mate-
However, while businesses and society strive to get advantages from the ongoing
rapid digitalization, it comes with several side-effects. Figures from the latest Global
E-Waste Monitor  indicate that digitalization currently accelerates the mass pro-
duction of E-Waste and will speed up more in the future. E-Waste is the fastest growing
waste stream in the world, with an annual growth rate of 3 to 4%. From 2014 to 2019,
it grew by 21%. Nonetheless, only 17.4% of global E-Waste was officially documented
and properly recycled in 2019. On one hand, this comes up with several negative envi-
ronmental and social impacts, not only at the end-of-life-phase of those products but
along the whole value chain.
A closer look at the evolution of the production and use of digital devices, such as;
smartphones, which we investigate in deep, support our argumentation. Smartphones
play a vital role in our daily life. People and businesses use them for communication,
shopping, navigation, entertainment, and many other activities with few screen touches.
The continuous consumption of smartphones contributes to a scarcity of non-renewable
resources since smartphone manufacturers use Rare Earth Element (REE) and other
precious metals. According to , only about 1% of smartphones are recycled, and
one reason behind this extremely low-rate is the technological complexity to recycle
REE. On the other hand, the raw material value of E-Waste offers vast economic op-
portunities. It is estimated  to be 5100 tons of smartphone content of precious and
critical metals in units put on the market by 2035 comparing to 1500 tons by 2020.
A periodic table that demonstrates the scarcity of elements used in smartphones was
demonstrated in 2019 on the 150th anniversary of the creation of the original periodic
table . Modern smartphones contain more than 30 different elements, in which gold,
silver, and copper are used for wiring and lithium and cobalt for the battery, and other
REE, including yttrium, terbium, and dysprosium. Even though having fractions of
grams is considered endangered. Many concerns are raised because about 17 elements
needed to manufacture smartphones are finite, and the continuous depletion of these
resources is alarming due to limited supplies, lack of recycling, or the location in con-
flict zones. A study by Yale University , tried to find possible replacements. How-
ever, they found 12 metals and metalloids, namely rhenium, rhodium, lanthanum, eu-
ropium, dysprosium, thulium, ytterbium, yttrium, strontium, thallium, magnesium, and
manganese, have no replacement at all because the substitution will be inadequate and
will decrease the performance.
But how to make use of these resources with the help of digital technologies such as
AI? So far this is still an open question , but this is a prerequisite for smart(er)
smartphone recycling, which is a significant component of smarter E-Waste manage-
The remainder of the paper is organized as follows. First, we further elaborate on
our motivation and challenges to make AI an enabler to CE in terms of E-Waste Man-
agement, we present the state of the art of automated waste management, and to narrow
our focus on smartphone recycling. Second, we present a TL method to classify
smartphones based on feature extraction. Third, our implementation of the TL is de-
scribed in detail, followed by demonstrating our experimental results and discussion of
optimizing the classification performance. Finally, we draw our conclusion and future
1.2 State of the Art of Automated Waste Management & Smartphone
Traditional waste recycling has many drawbacks: It uses intense manual labor leading
to high operation costs, and workers are exposed to these harmful substances through
inhalation, skin contact, or ingestion . Moreover, many industrial and household
appliances contain hazardous toxic materials like mercury that damages the human
Digital technologies could enhance waste management. It could do so not only the
end-of-life-phase of products but it could also extend their life-time and enhance their
product-life-cycle. To overcome these barriers and to gain CE benefits, many waste
management companies now understand the increasing need for smart Waste Manage-
ment Systems (WMS) and the automated disassembly of products to maintain sustain-
ability or stimulate eco-design products. Digital solutions are increasingly used to meet
the requirements of processing massive waste streams, e.g. identifying waste container
loads, tracking vehicle routes, etc. Real-time processing of a large volume of data with
the minimum human intervention will certainly support industrial decision-making.
Applying AI, including deep learning techniques, will enable building smart WMS.
This includes but is not limited to; E-Waste collection, recognizing waste patterns, sort-
ing and evaluating the material status, and estimating the behaviors of waste generators,
thereafter to support CE. All in all, we think that AI-enhanced E-Waste Management
will contribute to the development of CSC.
Smartphones are a specific type of E-Waste and there is also potential, but also a need
for further research on smart E-Waste management in this area. This is indicated by the
fact that the above-mentioned challenges drive leading smartphone manufacturers (Ap-
ple, Samsung, and Huawei) to take further measures to adopt a closed-loop system and
assess design sustainability, hence to develop and implement CE strategies.
Apple developed two disassembly robots, Liam, followed by Daisy, as a closed-loop
supply chain. The company announced that Daisy could recover all the materials like
Gold and REE used to manufacture its smartphones . Apple claimed that Daisy can
disassemble 15 different iPhone models at 200 devices per hour, which is more efficient
than any traditional recycling. They assemble devices by breaking down and separating
components to recover materials from iPhones. Daisy can disassemble 2 million de-
vices per year and recycle them automatically.
Samsung announced that the Re+ program has its sustainable promise to support CE.
According to , the company collected 3.55 million tons of end-of-life products be-
tween 2009 and 2018 through this program. It stated that the material compositions of
smartphones are: plastic, aluminum, steel, copper, cobalt (the primary resource used in
batteries), and gold and other materials, with the percentage of 35.1%, 20.2%, 10.6%,
10.0%, 8.6%, 15.5% respectively. Their new vision is to allow the company to design
the devices to be easy to repair, disassemble, and recycle, which will expand the life
span of products and improve durability.
Huawei also takes part in supporting CE through its Green Action program. Its ser-
vice centers took back almost 60 tons of spare parts every month in 2019 and involved
its customers in a credit-based recycling program . Furthermore, hundreds of thou-
sands of smartphone batteries were replaced each month of 2019 through the battery
replacement program at a fixed price, and they improve their maintenance quality
through discounted repair programs and even the EMUI 10.1 system that improves the
file fragmentation to prevent phones from freezing up for 18 months. Eventually, the
customers can use the product longer with fewer resources in the long term.
These companies can make products from recycled or renewed materials only by
using their own product design knowledge as a core prerequisite of recycling. It is worth
mentioning that modular phones like ARA by Google, G5 by LG, the Dutch FairPhone,
or the German ShiftPhone are examples of modular smartphones. They are considered
as best-practice in sustainable design and durability. These phones are easily disassem-
bled, contain less hazardous substances, long time warranty (mostly five years) as well
a transparent cost-breakdown . Unfortunately, they fail to take a big market share
because of their high costs in relation to lower-technical feasibility compared with con-
2 Method: Transfer Learning Approach - Extraction of
Information based on Visual Features
While describing the potentials of AI for smart E-Waste Management is easy, the de-
velopment of the respective solutions is a rather sophisticated task. Concerning the
technical challenges that face AI solutions, building an entire Neural Network (NN) is
a challenge even to AI experts. Therefore, rather than reinventing the wheel, we used
AlexNet  as a pre-trained model on a large-scale dataset, fine-tuned the model on a
new, relatively small training-set of smartphone images, and transferred the learned
characteristics to classify smartphones.
Challenges for smartphone classification emerge as their designs look similar re-
cently in terms of shape and size, especially when keypads, big antennas, buttons,
screen flips, and slides are abdicated. Instead, big touchscreens, all-glass front, multi-
cameras, and adjusted size to fit in hands became the typical design, in order to satisfy
The extraction of information based on visual features is often solved based on NN
. Convolutional Neural Networks (CNN) application has significant success in ob-
ject recognition and classification . Therefore, our method is designed to extract
information based on visual features.
2.1 Transfer Learning Method
It is labor-intensive to train NN from scratch because a huge data set is needed. Alter-
natively, an approach like TL could help to solve classification problems, e.g. different
smartphone models. Bear in mind that TL is considered as a supplement but not a re-
placement to learning techniques. To successfully implement TL, why, how, and when
to transfer should be clear beforehand.
Why Transfer Learning
In AI, new knowledge could be obtained by starting from scratch, but it needs a tre-
mendous amount of training data. The TL technique has verified its efficacy against the
scratch method’s training to tackle this problem. TL is a relatively new topic in the AI
domain. It is used when the source and target datasets have different features, and it
works efficiently when the target dataset has a small amount of data. The main concept
is to reuse specific parts of source samples into target samples to improve the attained
learning in a new task. Thus, our method is based on extracting features using a TL
approach that seeks good feature representation in the source and leads to better
smartphone classification accuracy and less error. Later in the implementation, we will
test the advantages of TL.
How to transfer and why AlexNet is used?
Image classification is one domain area in the field of deep learning . Using TL
techniques (Fine-tuning AlexNet, specifically) have impressive success in many fields
that underpin modern AI-enabled technology, to name but a few; biometrics  , med-
ical images , fault diagnosis in the industry , natural language processing .
However, smartphone classification received less attention.
Performing TL means choosing a pre-trained model that leverages the required task
as a starting point and then fine-tune it to achieve the desired results. AlexNet has been
used intensively in many applications as a leading model that uses TL for the following
•First, it is considered a deep NN because it has many hidden layers of non-linear
feature extractors, as we will describe them further in the network structure section.
•Second, it outperformed the other Non-deep learning method in the ImageNet Large
Scale Visual Recognition Challenge (ILSVRC) in 2012 .
•Third, it has a high-performance trade-off between accuracy and speed, thanks to
Rectified Linear Units (ReLU) that accelerates the convergence of the NN than using
saturation function like Tanh or Sigmoid .
Therefore, we used AlexNet in our approach, and we will describe the architecture in
When to transfer?
Even though TL has superior benefits, it is not merely a plug-and-play model. To decide
what features are maintained in the network is an open challenge. The pre-trained model
should be well understood before proceeding with any modifications.
3.1 Classifying Smartphones
In the implementation, we pass the training data to the network, and setting the options
of the training algorithm; then, we will train the network and optimize the performance.
Fig. 2, shows the system flowchart of the total implementation. The computing envi-
ronment was Matlab since it has a suitable deep learning toolbox, which allows us to
comprehensively customize solutions by creating, editing, visualizing, and analyzing
the CNN, on a core i5 Intel laptop with 16 GB RAM. An Allied GigE camera is used
for real-time testing.
We used the TL concept to classify 14 models of smartphones from different brands.
We start by building our dataset; then, we fine-tune the traditional AlexNet structure to
fit with our target output. Next, we set the training options to trigger the early stop.
After training the network, we monitor the performance, and we suggest to perform
controlling the error rate and data augmentation to enhance the generalization capabil-
ities. A technical description of the procedure is delivered in the following section.
Fig. 2. The system workflow of implementing smartphone classification with TL
In this paper, we suggest a fine-tuning of the pre-trained model of AlexNet. First, the
standard AlexNet is analyzed here. It has eight learned layers, as follows:
•Five convolutional layers (conv1 – conv5), which are basically used to extract fea-
tures. The information extracted from (conv1-conv3) represents the generic features
with different colors, texture, and intensity. Whereas, the next layers (conv4 –
conv5) extract the more refined features (or local patterns) like those with different
sizes and shapes.
•Three pooling layers, usually to downsample the features to implement faster com-
•Three Fully Connected (FC) layers: (FC6 – FC7) who are mainly used for features
that are more task-specific and prevent the model from overfitting while training,
(FC8) combines the previous features to present the output 1000 labels.
AlexNet is a large CNN that has successfully classified 1.2 million images with 1000
object labels, so this abundant data is rich with a wide variety of feature representations.
In the original pre-trained AlexNet architecture, the last third layer is configured to map
the extracted features from the previous layers to 1000 output classes; then, the softmax
layer acts as a normalization step to turn the raw values of the 1000 classes into a prob-
ability distribution of the image belongs to that class; thus, the sum of all elements in
this vector is equal to 1. Finally, the last layer takes the most probability and returns the
most likely class as a network output. We propose a network modification by freezing
the last three layers, replace them with (an FC layer, a softmax layer, a classification
output layer) to suit the new training-set, then retrain them, as illustrated in Fig. 3.
We control the behavior of the training algorithms to gain better training performance.
We split the dataset as 80% (320 images) for the training-set and 20% (80 images) for
the validation. We used the Stochastic Gradient Descent with Momentum (SGDM)
method as a training algorithm because it converges faster towards lower minima, and
it oscillates less. We set the mini-batch size to 20 and we found that the accuracy and
loss factor stabilize when the max epoch is equal to 20, where in each iteration one
mini-batch is trained and the number of epochs represents the number of times that the
network sees the entire dataset. We control the early stop when the validation error no
more improves to set a trade-off between the training time and accuracy. Following the
training, we evaluate the network performance using the validation-set during training.
It is an important step to check overfitting
3.2 Training the Network
After preparing the three previous components, we are ready to train our network. We
demonstrate different metrics to evaluate the classification efficiency; accuracy, and
loss function. Besides that, the confusion matrix of validation testing and real-time test-
ing will be conducted later to test the model performance. The accuracy represents the
percentage of the correctly classified trained images during an iteration to the number
of the entire dataset, which calculates the Root-Mean-Squared-Error (RMSE) in the
model gradients function. The error between the predictions and the true known class
is called the loss function. It defines the extent to which the actual outputs are correctly
predicted; practically, it represents the mini-batch loss. In the NN we aim to minimize
the loss function (see Fig. 4).
Fig. 3. Transfer learning approach by fine-tuning AlexNet structure
Fig. 4. The network performance before improvement
4 Results and Discussion
After training the network, we found that the validation accuracy is equal to 86.4%, and
it is stabilizing to be less than the training accuracy, which is not adequate. We recom-
mend the following steps to modify some training options to gain a better performance.
4.1 Controlling the Learning Rate
Choosing the learning rate is one of the challenging tasks in learning a CNN. In our
method, we schedule the learning rate by reducing updating the weights by slowing
down the learning rate initially to maintain the useful features, but then we speed up the
learning features. We set the dropout factor as 0.5 to obtain maximum regularization
. We found that the validation accuracy is 88.7%, but the model is underfitting (Fig.
4.2 Data Augmentation
Data augmentation is an automatically pre-processing stage during the training phase,
to cope with the imperfect images in terms of different angles, substandard lightings,
or not well-cropped or framed. This, in turn, prevents the overfitting problem by show-
ing the network, different variations of the same image, such as; rotation, reflection,
translation, shear, and scaling during the training phase. Subsequently, it leads to ef-
fortless adding multiple viewpoints of the same class of the non-altered data-set hence,
teaching the network that minor shifting, mirroring, or cropping of images does not
affect the prediction, but enhancing the classification accuracy. Consequently, it solves
the problem of having a few training data.
In our method, we use AlexNet that expects the input images’ size as 227x227x3, so
the training-set should be first resized to feed the first layer. Besides that, additional
randomly vertically flipping and vertically and horizontally translating the images are
performed to prevent the model from memorizing the training-set.
We perform reflection and translation on the X and Y axis, so our dataset was aug-
mented by 4 leading to 1680 images. We also shuffle the data before each epoch to
avoid discarding it every epoch. We found that the model is generalized, but the acti-
vation accuracy is 86.25% (Fig. 6).
Previously, we found that applying data augmentation or having a constant learning
rate leads to non-adequate network performance. We found that the model generalized
well without over or underfit, and the accuracy is enhanced to become 96.25% by
scheduling the learning rate, and we augment the dataset, as illustrated in Fig. 7. By
testing the (80 images) in the validation-set, a confusion matrix is demonstrated in Fig.
8. It is a numeric matrix that is used to measure the performance of the network by
creating a matrix from the true class and the predicted class. It shows how many obser-
vations in every cell, where the diagonal of the matrix shows the correctly classified
Fig. 5. The network performance with controlled learning rate
Fig. 6. The network performance with data augmentation
Fig. 7. The network performance including improvements
The normalized row and column (on the side of the matrix) display the percentage
of correctly classified class (highlighted in blue color) and the incorrectly classified
class (highlighted in orange color). We found that most of the smartphones are correctly
classified since the activation accuracy reached to 96.25%.
Apparently, from the confusion matrix we can calculate the loss function of the val-
idation-set, as the following equation:
Error rate of the ValidationSet = (The number of incorrectly classified objects in the
validationSet) / (The total number of validationSet) (1)
This means that the error rate here is equal to 0.0375 (3/80), which is very acceptable.
It also confirms the loss function value that is shown in Fig. 7.
Fig. 8. Confusion matrix of the validation set
4.3 Real-time Smartphone Classification
By using the real-testing set, illustrated in Fig. 2, we conducted a real-time smartphone
classification, by using the Allied GigE Camera and four examples of smartphone mod-
els. Fig. 9, shows that a high testing accuracy has been achieved based on visual features
only, with our proposed TL approach.
We found that the model leads to considerable results. Furthermore, this confirmed
our investigation that the TL does not require a massive dataset to get high accuracy,
even though the dataset is small. Besides that, TL is far easier than building the network
from scratch, and the training time is greatly reduced.
The results show that despite having no information about the smartphone design,
the model achieves good feasibility of the smartphone classification based on feature
fusion by using a TL technique.
Fig. 9. Real testing on four smartphone models
In this paper, we present how AI could support automated smartphone recycling, hence,
act as an enabler for CSC. We investigate a feature-based extraction of smartphones to
support CE. Currently, smartphone manufacturers start to endeavor to recycle their own
products, however, their recycling programs are designed to fit their own products only,
which may limit high recycling quotes. Therefore, we develop a feature-based TL ap-
proach that works without having any information about the design of the products. We
use the TL technique, by choosing AlexNet as a pre-trained model, to perform our test,
and to gain the advantages of TL techniques, as easier and faster way than training the
NN from scratch, which we prove in our results.
In consequence, we conclude that AI and CE could conjointly be applied to achieve
smart sustainability successfully. As we find that AI can help in transforming the E-
Waste management infrastructure into a closed-loop system, we conclude that AI can
pave the way towards CSC.
However, further research is needed. Smartphone recognition still faces more chal-
lenges even with state-of-the-art image classification methods, especially for the recent
smartphone models due to the high similarity in visual characteristics.
Future research will address these shortcomings. We suggest conducting non-de-
structive testing outside the visible light to detect the internal smartphone components,
e.g. the battery, camera, ID sensors, that helps in material recognition, by using a fusion
of sensors in different wavelengths to support automated recycling, hence the CE.
Last but not least, we argue that a fully-sustainable system would require rethinking
and changing behaviors of customers and smartphone manufacturers, respectively. This
would include, for instance, avoiding the replacement of smartphones every couple of
years unless they need maintenance and thinking in maintaining raw materials needed
by eco-design of future products.
 V. Albino, U. Berardi, and R. M. Dangelico, “Smart Cities: Definitions, Dimen-
sions, Performance, and Initiatives,” Journal of Urban Technology, vol. 22, no.
1, pp. 3–21, 2015, doi: 10.1080/10630732.2014.942092.
 M. de Jong, S. Joss, D. Schraven, C. Zhan, and M. Weijnen, “Sustainable–
smart–resilient–low carbon–eco–knowledge cities; making sense of a multitude
of concepts promoting sustainable urbanization,” Journal of Cleaner Produc-
tion, vol. 109, pp. 25–38, 2015, doi: 10.1016/j.jclepro.2015.02.004.
 R. P. Dameri, Ed., Smart City Implementation: Creating Economic and Public
Value in Innovative Urban Systems. Cham, s.l.: Springer International Publish-
ing, 2017. [Online]. Available: http://dx.doi.org/10.1007/978-3-319-45766-6
 S. Prendeville, E. Cherim, and N. Bocken, “Circular Cities: Mapping Six Cities
in Transition,” Environmental Innovation and Societal Transitions, 2017, doi:
 M. Antikainen, T. Uusitalo, and P. Kivikytö-Reponen, “Digitalisation as an En-
abler of Circular Economy,” Procedia CIRP, vol. 73, pp. 45–49, 2018, doi:
 S. Ritzén and G. Ö. Sandström, “Barriers to the Circular Economy – Integration
of Perspectives and Domains,” Procedia CIRP, vol. 64, pp. 7–12, 2017, doi:
 J. Kirchherr, D. Reike, and M. Hekkert, “Conceptualizing the circular economy:
An analysis of 114 definitions,” Resources, Conservation and Recycling, vol.
127, pp. 221–232, 2017, doi: 10.1016/j.resconrec.2017.09.005.
 Thalesgroup, Facial recognition in 2020 (7 trends to watch). [Online]. Availa-
ernment/biometrics/facial-recognition (accessed: Oct. 4 2020).
 S. Hommel, M. A. Grimm, D. Malysiak, and U. Handmann, “APFel - fast multi
camera people tracking at airports, based on decentralized video indexing,” Sci-
ence2 - Safety and Security, HOMELAND SECURITY UG, Hemer, Germany,
2: 48–55, 2014.
 Zengeler, N. and Arntz, A. and Keßler, D. and Grimm, M. and Qasem, Z. and
Jansen, M. and Eimler, S. and Handmann, U., “Person Tracking in Heavy Indus-
try Environments with Camera Images,” in 10th EAI International Conference
on Sensor Systems and Software, Braga, Portugal, 2020, pp. 324–336.
 M. A. Schreurs and S. D. Steuwer, Autonomous Driving - Political, Legal, So-
cial, and Sustainability Dimensions: Springer Berlin Heidelberg, 2015.
 P. Mallozzi, P. Pelliccione, A. Knauss, C. Berger, and N. Mohammadiha, “Au-
tonomous Vehicles: State of the Art, Future Trends, and Challenges,” in Auto-
motive Systems and Software Engineering, Y. Dajsuren and M. van den Brand,
Eds., Cham: Springer International Publishing, 2019, pp. 347–367.
 U. Handmann, T. Kalinke, C. Tzomakas, M. Werner, and W.v. Seelen, “An im-
age processing system for driver assistance,” Image and Vision Computing, vol.
18, no. 5, pp. 367–376, 2000, doi: 10.1016/S0262-8856(99)00032-3.
 A. Rabie and U. Handmann, Proceeding of the 11th World Congress on Intelli-
gent Control and Automation: NFC-based person-specific assisting system in
home environment: IEEE, Jun. 2014 - Jul. 2014.
 T. Kopinski, F. Sachara, and U. Handmann, Proceedings of the 13th Interna-
tional Conference on Mobile and Ubiquitous Systems: Computing, Networking
and Services - MOBIQUITOUS 2016: A Deep Learning Approach to Mid-air
Gesture Interaction for Mobile Devices from Time-of-Flight Data. New York,
New York, USA: ACM Press, 2016.
 C. Nieß, J. Fey, D. Schwahlen, M. Reimann, and U. Handmann, Applying step
heating thermography to wind turbine rotor blades as a non-destructive testing
method. Telford, UK, 2017.
 J. Fey, C. Djahan, T. A. Mpouma, J. Neh-Awah, and U. Handmann, 2017 Far
East NDT New Technology & Application Forum (FENDT): Active Thermo-
graphic Structural Feature Inspection of Wind-Turbine Rotor: IEEE, Jun. 2017
 ITU, “End of life management for ICT equipment,” 2012. [Online]. Available:
 C. P. Baldé, Forti V., V. Gray, R. Kuehr, and P. Stegmann, The Global E-waste
Monitor 2017: Quantities, Flows, and Resources. Bonn/Geneva/Vienna: United
Nations University (UNU); International Telecommunication Union (ITU); In-
ternational Solid Waste Association (ISWA), 2017. Accessed: Dec. 14 2017.
 N. Gurita, M. Fröhling, and J. Bongaerts, “Assessing potentials for mo-
bile/smartphone reuse/remanufacture and recycling in Germany for a closed
loop of secondary precious and critical metals,” Jnl Remanufactur, vol. 8, 1-2,
pp. 1–22, 2018, doi: 10.1007/s13243-018-0042-1.
 A. Pagoropoulos, D. C.A. Pigosso, and T. C. McAloone, “The Emergent Role of
Digital Technologies in the Circular Economy: A Review,” Procedia CIRP, vol.
64, pp. 19–24, 2017, doi: 10.1016/j.procir.2017.02.047.
 A. Julander et al., “Formal recycling of e-waste leads to increased exposure to
toxic metals: an occupational exposure study from Sweden,” Environment inter-
national, vol. 73, pp. 243–251, 2014, doi: 10.1016/j.envint.2014.07.006.
 Ellen MacArthur Foundation (EMF) and Google, Eds., “Artificial intelligence
and the circular economy,” Jul. 2020. Accessed: Jul. 17 2020. [Online]. Availa-
 A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet classification with
deep convolutional neural networks,” 2017, doi: 10.1145/3065386.
 T. Kopinski, F. Sachara, and U. Handmann, “A Deep Learning Approach to
Mid-air Gesture Interaction for Mobile Devices from Time-of-Flight Data,” in
Proceedings of the 13th International Conference on Mobile and Ubiquitous
Systems: Computing, Networking and Services - MOBIQUITOUS 2016: A Deep
Learning Approach to Mid-air Gesture Interaction for Mobile Devices from
Time-of-Flight Data, Hiroshima, Japan, 2016, pp. 1–9.
 A. A. Almisreb, N. Jamil, and N. M. Din, “Utilizing AlexNet Deep Transfer
Learning for Ear Recognition,” in 2018 Fourth International Conference on In-
formation Retrieval and Knowledge Management (CAMP), Kota Kinabalu, Ma-
laysia, Mar. 2018 - Mar. 2018, pp. 1–5.
 H.-C. Shin et al., “Deep Convolutional Neural Networks for Computer-Aided
Detection: CNN Architectures, Dataset Characteristics and Transfer Learning,”
IEEE transactions on medical imaging, vol. 35, no. 5, pp. 1285–1298, 2016,
 X. Shi, Y. Cheng, B. Zhang, and H. Zhang, “Intelligent fault diagnosis of bear-
ings based on feature model and Alexnet neural network,” in 2020 IEEE Inter-
national Conference on Prognostics and Health Management (ICPHM), De-
troit, MI, USA, Jun. 2020 - Jun. 2020, pp. 1–6.
 R. Almodfer, S. Xiong, M. Mudhsh, and P. Duan, “Enhancing AlexNet for Ara-
bic Handwritten words Recognition Using Incremental Dropout,” in 2017 IEEE
29th International Conference on Tools with Artificial Intelligence (ICTAI), Bos-
ton, MA, Nov. 2017 - Nov. 2017, pp. 663–669.
 E. Uçar, M.-A. Le Dain, and I. Joly, “Digital Technologies in Circular Economy
Transition: Evidence from Case Studies,” Procedia CIRP, vol. 90, pp. 133–136,
2020, doi: 10.1016/j.procir.2020.01.058.
 I. Barletta, B. Johansson, K. Cullbrand, M. Bjorkman, and J. Reimers, “Foster-
ing sustainable electronic waste management through intelligent sorting equip-
ment,” in 2015 IEEE International Conference on Automation Science and En-
gineering (CASE), Gothenburg, Sweden, Aug. 2015 - Aug. 2015, pp. 459–461.
 AMP Robotics, AMP Neuron — AMP Robotics. [Online]. Available: https://
www.amprobotics.com/amp-neuron (accessed: Oct. 1 2020).
 A. Nygaard, “Specific Investments in Closed Loop-Technology instead of
“Blood Metals”,” 2019.
 EuChemS, Element Scarcity - EuChemS Periodic Table - EuChemS. [Online].
(accessed: Oct. 3 2020).
 YaleNews, For metals of the smartphone age, no Plan B. [Online]. Available:
Oct. 1 2020).
 Apple Newsroom, Apple expands global recycling programs. [Online]. Availa-
ble: https://www.apple.com/newsroom/2019/04/apple-expands-global-r ecy-
cling-programs/ (accessed: Oct. 1 2020).
 Samsung levant, Resource Efficiency | Environment | Sustainability | Samsung
LEVANT. [Online]. Available: https://www.samsung.com/levant/aboutsamsung/
sustainability/environment/resource-efficiency/ (accessed: Oct. 1 2020).
 huawei, Green Pipe - Huawei Sustainability. [Online]. Available: https://
cessed: Oct. 1 2020).
 M. Proske, K. Schischke, P. Sommer, T. Trinks, N. F. Nissen, and K.-D. Lang,
“Experts View on the Sustainability of the Fairphone 2,” in 2016 Electronics
Goes Green 2016+ (EGG), Berlin, Sep. 2016 - Sep. 2016, pp. 1–7.
 U. Handmann, G. Lorenz, T. Schnitger, and W. von Seelen, Fusion of Different
Sensors and Algorithms for Segmentation. Stuttgart, Germany: IEEE Interna-
tional Conference on Intelligent Vehicles ‘98, 1998.
 F. Sachara, T. Kopinski, A. Gepperth, and U. Handmann, “Free-hand gesture
recognition with 3D-CNNs for in-car infotainment control in real-time,” in 2017
IEEE 20th International Conference on Intelligent Transportation Systems
(ITSC): Free-hand Gesture Recognition with 3D-CNNs for In-car Infotainment
Control in Real-time, Yokohama, Oct. 2017 - Oct. 2017, pp. 959–964.
 K. Chatfield, K. Simonyan, A. Vedaldi, and A. Zisserman, “Return of the Devil
in the Details: Delving Deep into Convolutional Nets,” 2014. [Online]. Availa-
 D. P. Kingma, T. Salimans, and M. Welling, “Variational Dropout and the Local
Reparameterization Trick,” 2015. [Online]. Available: https://arxiv.org/pdf/
 S. Wiegand, C. Igel, and U. Handmann. Evolutionary optimization of neural
networks for face detection. In ESANN 2004, 12th European Symposium on Ar-
tificial Neural Networks, Bruges, Belgium, Proceedings, pages 139–144, 2004.
 S. Wiegand, C. Igel, and U. Handmann. Evolutionary multi-objective optimiza-
tion of neural networks for face detection. International Journal of Computa-
tional Intelligence and Applications, 4(3):237–253, 2004.