______________________________________________________PROCEEDING OF THE 31ST CONFERENCE OF FRUCT ASSOCIATION

TrashBox: Trash Detection and Classification using

Quantum Transfer Learning
Nikhil Venkat Kumsetty, Amith Bhat Nekkare, Sowmya Kamath S., Anand Kumar M.
National Institute of Technology Karnataka, Surathkal, India
nikhilvenkat26@gmail.com, amithbhat01@gmail.com, sowmya.kamath@nitk.edu.in, m anandkumar@nitk.edu.in

Abstract—The problem of effective disposal of the trash a scarce commodity, it is imperative to reduce the space taken
generated by people has rightfully attracted major interest by landfills.
from various sections of society in recent times. Recently, deep Furthermore, the advent of mass manufacturing has seen the
learning solutions have been proposed to design automated
mechanisms to segregate waste. However, most datasets used proliferation of a wide variety of product classes, ranging from
for this purpose are not adequate. In this paper, we introduce electronics to medical products. For example, there has been a
a new dataset, TrashBox, containing 17,785 images across seven staggering growth in the number of electronic items produced
different classes, including medical and e-waste classes which in the last 20 years, and therefore, similar rise in the e-waste
are not included in any other existing dataset. To the best of our handled by waste processing plants. E-waste is one of the most
knowledge, TrashBox is the most comprehensive dataset in this
field. We also experiment with transfer learning based models recyclable class of trash items, for example, smartphones are
trained on TrashBox to evaluate its generalizability, and achieved made using rare earth elements, gold, and other valuable mate-
a remarkable accuracy of 98.47%. Furthermore, a novel deep rials, all of which can be salvaged without too much difficulty
learning framework leveraging quantum transfer learning was using standardized processes. Similarly, with the COVID-19
also explored. Experimental evaluation on benchmark datasets pandemic, there has been a sharp increase in the number of
has shown very promising results. Further, parallelization was
incorporated, which helped optimize the time taken to train the surgical masks, gloves, and other such items discarded as litter,
models, recording a 10.84% improvement in the performance often without any regard to mandated bio-hazardous material
and 27.4% decline in training time. disposal procedures. It is imperative that such hazardous items
be handled with appropriate care, to protect trash disposal
squads and other environmental contamination.
I. I NTRODUCTION To reduce the the amount of waste sent to landfills, there
is a need for automated systems for the segregation of the
The past decade has evidenced an explosive increase in the same. However, manual sorting is a tedious and often haz-
amount of trash generated every day by urban populations. ardous process, which does not scale well as being labour
However, the current practice of directly discarding away and cost-prohibitive. The trash sorting machines employed
items into the trashcan is highly unsustainable - particularly currently work on physical sorting processes, not intelligent
when raw materials are finite resources, and will eventually computerized methods, which can help in more accurate class-
be exhausted. Instead, such products could be reclaimed, wise segregation. Specifically, deep neural models trained on
processed and recycled, and thus be used again. Recent surveys a trash image dataset with large variety of classes, can help
have reported that, in Delhi, roughly 80% of the waste sent classify trash objects based on their type, and boost recyclable
to landfills daily could be recycled [1]. A similar argument product reclamation and productivity.
can be made from the perspective of reducing the waste In this paper, a novel approach leveraging quantum trans-
disposed off every day. Currently, the two major forms of fer models for trash classification is presented. For initial
waste disposal are incineration, where the object is simply experiments, we experiment with transfer learning models
burned, and landfilling, where the object is dumped in a pre- involving the use of models pre-trained on a very large datasets
determined spot. Of these two methods, landfilling can be (ImageNet, COCO, etc.). These pre-trained models are trained
said to be far more popular, due to the lack of machinery to solve a particular problem through a new large-scale dataset
or investment required to start operations. However, while encompassing multiple fine-grained classes, which are not
incineration, even in the most controlled circumstances, does supported by most existing trash classification datasets. As the
produce air pollution, landfills are much more harmful in model does not start off with random weights, instead using
nature. Landfills are known to pollute the groundwater in the weights inherited from the previously trained dataset. This
the region, due to toxic chemicals leaching into the water helps to increase the training accuracy, reduce the amount of
reservoirs. Fires are also known to break out spontaneously computations performed and increase the overall performance
in such sites, due to the presence of highly flammable items. of the model. Also, we adopted quantum computing for the
Landfills are also quite an eyesore, and give off an odious improvement and optimization of many processes involved for
smell as one approaches the area. Moreover, as land becomes reducing the time and resources spent on performing complex

ISSN 2305-7254

Authorized licensed use limited to: Chaitanya Bharathi Institute of Tech - HYDERABAD. Downloaded on August 16,2024 at 03:49:35 UTC from IEEE Xplore. Restrictions apply.
 ______________________________________________________PROCEEDING OF THE 31ST CONFERENCE OF FRUCT ASSOCIATION

computations in neural model layers. Further, they changed the input size of the Inception model
We focus on a specific field of quantum machine learn- to 299x299 as that is the default image dimension of the
ing called quantum transfer learning, thereby leveraging two Raspberry Pi camera, and reported an accuracy of 92.5%.
powerful concepts, transfer learning and quantum computing Masand et al. [7] created ScrapNet, a deep learning model
to find a possible solution for trash object classification. In adapted form the architecture of the EfficientNet transfer
addition, to alleviate the problem of inadequate waste datasets, learning model. They assessed the viability of this dataset
we introduce TrashBox, a new and comprehensive waste by testing it on the benchmark TrashNet dataset, where they
dataset with 17,785 real-world trash images divided into 7 achieved a 98% accuracy. They also created a new dataset
classes : cardboard,e-waste, glass, medical waste, metal, paper, consisting of 8,135 images, by collating images present in
and plastic. Hence, hitherto neglected classes such as medical various pre-existing datasets and standardizing them. Running
waste and e-waste are also incorporated. an EfficientNet B3 model on this dataset, they achieved an
The rest of this paper is organized as below: Section 2 accuracy of 92.87%. Shi et al. [8] proposed a waste clas-
presents a discussion on recent and notable work carried out sification method based on a multilayer hybrid convolution
in the field of trash classification and quantum computing, neural network. The structure of the network, while similar
specially quantum transfer learning. Section 3 discusses the de- to VGGNet, is simpler, with higher output performance and
tailed methodology, with respect to the details of the TrashBox fewer input parameters. Experiments with TrashNet achieved
dataset, and construction and training optimization of quantum a classification accuracy of up to 92.6%.
transfer learning models. In Section 4, the specifics of the In recent years, there has been remarkable progress in
experiments performed and the results obtained are discussed the field of quantum computing. In particular, it has been
in detail, followed by conclusion and future work. applied to a wide variety of tasks for achieving time and
cost optimizations when complex computations are to be
performed in neural model layers. Killoran et al. [9] in-
II. R ELATED W ORK troduced a general method of constructing neural networks
In this section, we present an analysis of the works related on quantum computers. The quantum neural network thus
to the area of interest. [2] provided a benchmark dataset, called built is a variational quantum circuit built in the continuous-
TrashNet, for trash detection and classification. TrashNet con- variable architecture. They performed extensive modelling
tains 2527 plain-background images from 6 different classes, experiments on the constructed network, such as, fraud de-
namely metal, plastic, paper, cardboard, glass and (other) trash. tection classifier, hybrid classical-quantum auto-encoder, etc.,
The authors experimented with CNN and SVM models and which demonstrated its wide-spread capabilities. Mari et al.
reported that the SVM model achieved the best accuracy at [10] proposed a framework of transfer learning consisting of
63%. Moreover, as the authors themselves note, the number hybrid computational models that were constructed using a
of data points in the dataset is very small, hence it is not mixture of variational quantum circuits and classical neural
very suitable for training neural models. Recently, another network models. In particular, 3 types of novel classical-
dataset, called TACO (Trash Annotations in Context) [3] for quantum models were proposed - CQ (classical-quantum),
trash detection and classification was proposed. The authors QC (quantum-classical), and QQ (quantum-quantum). They
put together a dataset containing “waste in the wild” trash also proposed the concept of dressed quantum circuits with
objects, i.e., images of trash objects and litter found in various which custom quantum circuits can be constructed based on
locations such as roads, woods, beaches, etc. The images in the requirement. All the proposed models were implemented
TACO are manually labeled and segmented according to a on quantum computers of IBM and Rigetti. Mogapalli et al.
hierarchical taxonomy to train and evaluate object detection [11] proposed a quantum transfer learning based approach for
algorithms. While the dataset is crowdsourced, it currently has different image classification tasks such as organic and recy-
only 1,500 annotated images, which is lower than TrashNet. clable classification of trash, TB detection from chest X-Rays,
Several works have explored the effectiveness of machine and detection of cracks in concrete structures. The authors
and deep learning models in the context of classifying waste use a concatenation of pre-trained classical feature extractor
objects. Adedeji et al. [4] used the ResNet-50 transfer learning with a quantum circuit as classifier. Various experiments were
model to classify trash objects, using the TrashNet dataset. performed using VGG-19, DenseNet-169 and AlexNet as pre-
To compensate for less frequency of images, the authors trained classifiers. However, no model was concluded to be
augmented the dataset by using various techniques such as better for classification, with different models outperforming
shearing and scaling. Finally, they use a multi-class SVM the others in different tasks.
model where the classification takes place. The authors got an Following the comprehensive review, we observed a glaring
accuracy of 87% with the TrashNet dataset with this model. lack of an adequate dataset with multiple categories for trash
Azis et al. [5] used a simple CNN model to classify trash classification research. The most commonly used dataset,
objects, using the Inception-v3 transfer learning model. The TrashNet, has only 2,527 trash images, which is hardly suffi-
training dataset used was the Plastic Model Detection dataset cient for a such a complex problem. Moreover, many relevant
[6] which contained 2,400 images. To simulate real-world and significant classes of trash were not covered by existing
conditions, the authors used Raspberry Pi as their processor. datasets. Chief among these were the medical waste and e-

---------------------------------------------------------------------------- 126 ----------------------------------------------------------------------------

Authorized licensed use limited to: Chaitanya Bharathi Institute of Tech - HYDERABAD. Downloaded on August 16,2024 at 03:49:35 UTC from IEEE Xplore. Restrictions apply.
 ______________________________________________________PROCEEDING OF THE 31ST CONFERENCE OF FRUCT ASSOCIATION

waste categories, which have received scant attention so far. a comprehensive search on the web. For this purpose, a
In this paper, we present a dataset with ample number of trash batch download software tool called WFdownloader [13] was
images per class. We have also included adequate number of used. This tool allowed us to download images in bulk from
images in the medical waste and e-waste classes. We also Google Images, Bing Images, among a variety of sources. The
propose the use of the quantum transfer-learning models to downloaded images were then manually cropped and sorted
perform the task of trash classification. into their respective classes. The details of the sub-classes are
given in Table I. The distribution of each class of the TrashBox
dataset is illustrated in Fig. 2.
III. P ROPOSED M ETHODOLOGY
This section detailed the proposed methodology designed
for creating the new multi-class trash dataset, that provides
large-scale images across multiple categories. Experiments
were performed using traditional transfer learning models to
validate their performance on the new dataset in comparison to
other standard datasets. Furthermore, the concept of quantum
transfer learning was also used to improve the model’s object
detection and feature extraction capabilities, and increased the
efficiency of the existing model by parallel processing within
the model. We describe the detailed workflow in Fig.1.

Fig. 2. Distribution of Trash images in TrashBox

The dataset can be downloaded from GitHub1 .

TABLE I. TRASHBOX DATASET - CLASS-WISE

STATISTICS

Trash Classes Sub-classes No. of images Total

Cardboard Assorted cardboard objects 2414 2414
E-waste Electrical chips 615
Laptops & smartphones 774 2883
Fig. 1. Workflow diagram of the proposed methodology Small appliances 926
Electric wires 568
Glass Assorted glass objects 2528 2528
A. Proposed Dataset - TrashBox Medical Syringes 507
waste Surgical gloves 496 2010
To address the lack of existing trash detection and classifi- Surgical masks 500
Medicines 507
cation datasets, a new dataset was put together. The detailed
Metal Beverage cans 1000
review revealed that the deficiencies in the various benchmark Construction scrap 539
datasets used in this field, such as TrashNet [2] and TACO Spray cans 500 2586
[3]. We observed that many benchmark datasets such as Metal containers 505
Miscellaneous metal 42
TrashNet lack the frequency of images required for the task Plastic Plastic bags 504
of classification of trash objects. In fact, this has been noted Plastic bottles 571
by the authors themselves in their paper [2] on the subject. Plastic containers 580 2669
Plastic cups 507
Further, two of the major modern day waste types are e-waste Cigarette butts 507
and medical waste. However, most datasets do not contain Paper Tetra pak 794
any images for these critical classes. This is important in e- News paper 200 2695
Paper cups 639
waste management [12] as implementing the right practices to Other paper objects 1062
recycle can help conserve non-renewable natural resources.
In order to remedy this issue, we created a dataset called
TrashBox that contains various trash objects in diverse en- B. Transfer Learning Models
vironments. The images in TrashBox were classified into 7
For the purpose of measuring performance and benchmark-
classes - medical waste, e-waste, glass, plastic, cardboard,
ing, we make use of popular standard transfer learning models.
paper, metal. Furthermore, these classes are divided into sub-
In particular, we use the ResNet-34, ResNet-50, ResNet-101,
classes to facilitate the distinction between various trash ob-
jects, and to enable further research in this field. TrashBox was 1 TrashBox dataset on GitHub, https://github.com/nikhilvenkatkumsetty/
prepared by extracting images of trash objects by performing TrashBox

---------------------------------------------------------------------------- 127 ----------------------------------------------------------------------------

Authorized licensed use limited to: Chaitanya Bharathi Institute of Tech - HYDERABAD. Downloaded on August 16,2024 at 03:49:35 UTC from IEEE Xplore. Restrictions apply.
 ______________________________________________________PROCEEDING OF THE 31ST CONFERENCE OF FRUCT ASSOCIATION

VGG-19, and DenseNet-121 models in various scenarios. We The quantum circuits described above are constructed us-
construct the models by importing pre-trained weights of the ing several layers. A quantum layer consisting of one qubit
ImageNet dataset. Then, the imported model is optimized by Hadamard gate that maps the basic states (0 and 1) to a qubit
adding a Global Average Pooling, Batch Normalization, and that exists in equally probable superposition of both the basic
Dense output layers to the model. The main goal of adding states. A quantum layer of parametrized qubit rotations [15]
the Global Average Pooling layer is to eliminate the need to around the x-axis. A quantum layer of parametrized qubit
fine-tune the hyper-parameters of the classical transfer learning rotations around the y-axis. A quantum entangling layer helps
models. We then start training the model on each dataset by to perform rotations on each qubit and then cascading the 2-
freezing all base layers and training the output layer by a large qubit entangling gates.
learning rate. To further improve the accuracy of the model,
we reduce the learning rate after a few epochs. The advantage D. Parallelization of QTL Model Training
of this approach is that it does not change the weights of the After construction, the QTL model was observed to take up
base layers drastically in the first few epochs when the last excessive amounts of computing resources and time during
few layers have not yet stabilized. the training phase, which we attribute to the models’ intricate
nature. To resolve this issue, we focused on designing a par-
C. Quantum Transfer Learning Models allelization strategy to speed up the model’s training process.
Towards this, we assigned a replica of the neural network to
In this section, we discuss our implementation of the quan- each thread. This helps to train the model in concurrent batches
tum transfer learning (QTL) methodology to classify trash ob- depending upon the number of threads and the computation
jects. In particular, we implement a hybrid classical-quantum power of the processing unit. Once the training process is
neural network, which helps to train the transfer learning complete, the weights computed are updated to their respective
models more effectively and process high-dimensional data threads. At the end, each batch of training data is averaged,
in an optimal way. This network is constructed by embedding thus generating the total performance of the model. We must
the quantum circuit layers into the classical transfer learning also be mindful about the effects of parallelizing a neural
model. As illustrated in Fig. 1, the quantum circuit layers are network on the time taken to finish the process and the
introduced towards the closing end of the classical transfer performance of the model. Since the TrashBox dataset has
learning model. This is done to: (a) encourage the feature large number of data-points to process, parallel computing
extraction capabilities of the classical transfer learning model helped to decrease the time taken to train the network. Detailed
and (b) help the quantum circuit layers process the input data observations regarding the various aspects of this strategy are
and improve the effectiveness of the classification task. presented in the following section.
The implemented QTL model mainly consists of two
components, the Preprocessing block and Quantum circuit
network. IV. E XPERIMENTAL R ESULTS AND A NALYSIS
A. Experimental Setup
1) Preprocessing block: To build the first part of the QTL
The experimental validation of the proposed models was
model, we import a pre-trained network trained on
performed on two standard datasets - TrashNet, TACO and the
ImageNet [14] dataset. After importing the network, we
custom dataset - TrashBox. The models were constructed using
remove the last fully connected layers that are used to
Python 3.10.1. The PyTorch open-source library was used to
perform the task of classification. Hence, we have a
develop the deep neural networks, and PennyLane [16] cross-
neural network that acts as a pre-processing block that
platform library to construct the quantum neural networks. The
extracts features from the input data. We then merge
models were trained on Nvidia P100 and T4 GPUs with a
this modified pre-trained network with the custom-built
minimum of 16 GB RAM, and all the proposed models were
dressed quantum circuit network, which help to con-
run for 100 epochs during this experimentation.
struct custom quantum circuits based on the requirement
of our model. B. Evaluation of the proposed models
2) Quantum circuit network: The second part is built using We measured the performance of TrashBox by using it
an amalgamation of various quantum layers. In this to train state-of-the-art deep neural models like ResNet-34,
case, an embedding layer is used to initialize the states ResNet-50, ResNet-101, DenseNet-121 and VGG-19. Table II
of quantum bits and input the features extracted from shows the results obtained by the classical transfer learning
the classical transfer learning model, so as to further models for the TrashBox dataset. We observed that the ResNet-
process the data and classify the trash objects. Then, 101 model achieved the best results among the considered
the variational layers perform the processing of the data models for this classification task. This model was able to
input into the embedding layer. Finally, the measurement classify the trash objects with a training accuracy of 98.86%,
layer generates a classical vector as output, by keeping validation accuracy of 98.29% and testing accuracy of 98.47%.
track of expectation value of the Pauli-Z operator for Next, we performed experiments on our quantum transfer
each qubit in the quantum layers. learning models. We created a train, test, validation split for

---------------------------------------------------------------------------- 128 ----------------------------------------------------------------------------

Authorized licensed use limited to: Chaitanya Bharathi Institute of Tech - HYDERABAD. Downloaded on August 16,2024 at 03:49:35 UTC from IEEE Xplore. Restrictions apply.
 ______________________________________________________PROCEEDING OF THE 31ST CONFERENCE OF FRUCT ASSOCIATION

TABLE II. PERFORMANCE OF CLASSICAL TRANSFER LEARNING MODELS model, which results in the building of a lighter but deeper
TRAINED ON TRASHBOX network in each layer. This in turn helps reduce the number of
computation operations performed in each layer and preserves
Model Training Validation Testing the weights of the inherited ImageNet dataset for a longer
ResNet-34 92.17% 92.13% 92.20% period, thereby improving the performance at a faster rate than
ResNet-50 96.44% 94.83% 93.36% the VGG model.
ResNet-101 98.86% 98.29% 98.47%
DenseNet-121 95.47% 93.60% 93.94% C. Evaluation of training the proposed models in parallel
VGG-19 96.91% 96.08% 95.63%
To verify the effect of parallelization of the training process
of the QTL models, additional experiments were conducted.
TABLE III We implemented parallel programming on the neural networks
P ERFORMANCE OF QTL MODELS TRAINED ON VARIOUS DATASETS
in order to improve the performance of the models and to
Dataset Model Training Validation Test
reduce the amount of time it takes to train a QTL model. This
TrashNet Quantum ResNet-34 76.97% 77.45% 77.56% step was important especially for QTL model since, the GPU
Quantum ResNet-50 81.10% 80.20% 80.49% were not very capable of handling such large computations as
Quantum ResNet-101 79.54% 78.97% 77.34% they had limited processing power. Moreover, the variational
Quantum DenseNet-121 78.52% 78.50% 78.1% quantum networks we used were designed to be processed
Quantum VGG-19 76.25% 77.50% 77.43% with Quantum Processing Units (QPU).
TACO Quantum ResNet-34 78.83% 79.20% 79.5% Firstly, we investigated the effects of varying the number
Quantum ResNet-50 82.30% 81.25% 82.1% of threads to the time required to complete the training of the
Quantum ResNet-101 80.14% 80.74% 80.34%
model. From Fig. 3, it can be seen that, the time required
Quantum DenseNet-121 79.22% 79.13% 79.03%
to train the models decreases with increase in the number
Quantum VGG-19 78.40% 78.13% 78.11%
TrashBox Quantum ResNet-34 81.53% 80.16% 80.38%
of threads, as expected. However, we note that the rate of
Quantum ResNet-50 85.38% 85.12% 84.97% decrease is higher for the QTL model than the classical transfer
Quantum ResNet-101 83.65% 82.14% 82.94% learning model. In particular, while the quantum model and
Quantum DenseNet-121 81.66% 81.27% 81.9% classical model take 11.5 hours and 10.5 hours respectively to
Quantum VGG-19 80.22% 79.56% 79.33% train serially, they both take 9.25 hours to train on 5 threads,
and 8.25 and 8.75 hours respectively to train on 7 threads.

analyzing the ResNet-34, ResNet-50, ResNet-101, DenseNet-

121, and VGG-16 QTL models. Table III shows the results
obtained by the QTL models run on the augmented TrashNet
dataset. We observed that the Quantum ResNet-50 model
achieved the best results among the considered models for this
analysis. This model was able to classify the trash objects with
a training accuracy of 81.1%, validation accuracy of 80.2%
and testing accuracy of 80.49%. It achieved best results for
the TACO and TrashBox datasets as well, with a training
accuracy of 82.3%, validation accuracy of 81.25% and testing
accuracy of 82.1% for TACO, and a training accuracy of
85.38%, validation accuracy of 85.12% and testing accuracy Fig. 3. Number of Threads vs Training time
of 84.97%. for TrashBox.
From Table III, it can be observed that the Quan- We also investigated the effects of varying the number of
tum ResNet-50 model trained on all datasets marginally out- threads to the accuracy of the trained model. From Fig. 4,
performed the Quantum ResNet-101 model. This may be it can be observed that, while the performance of the neural
because, ResNet-101, DenseNet-121, and other similar mod- networks improves upon parallelizing, it does so only until a
els are complex models which require more computational certain no. of threads, specifically, for a thread range of 6 to
resources and more time to train. Hence, these models are 8 threads. Beyond this, the performance degrades. Hence, we
more prone to overfit to the train data. However, we avoid this implemented our QTL model for 7 threads. As we increase
problem by optimizing the model to self-adjust its parameters the no. of threads, the neural networks ability to retain and
based on the results obtained by processing the validation uniformly update the parameter weights during the training
set. Therefore, it gives a slightly lower performance than the of each batch will be impacted severely, thus, resulting in a
ResNet-50 model. It can also be observed that the Quan- decrease in the accuracy of the trained models. Therefore,
tum ResNet-34, Quantum ResNet-50, and Quantum ResNet- we achieve 10.84% improvement in the training accuracy
101 models achieved better results than the Quantum VGG- and 27.4% decline in the time taken to train the models by
19 model. This is because of the architecture of the ResNet parallelizing the QTL model.

---------------------------------------------------------------------------- 129 ----------------------------------------------------------------------------

Authorized licensed use limited to: Chaitanya Bharathi Institute of Tech - HYDERABAD. Downloaded on August 16,2024 at 03:49:35 UTC from IEEE Xplore. Restrictions apply.
 ______________________________________________________PROCEEDING OF THE 31ST CONFERENCE OF FRUCT ASSOCIATION

under development.
R EFERENCES
[1] S. Lahiry. (2017) Percentage of land-filled waste recyclable in
delhi. [Online]. Available: https://www.downtoearth.org.in/blog/waste/
indias-challenges-in-waste-management-56753
[2] G. Thung and M. Yang, “Classification of trash for recyclability status,”
2016. [Online]. Available: https://github.com/garythung/trashnet
[3] P. F. Proença and P. Simões, “Taco: Trash annotations in context for
litter detection,” arXiv preprint arXiv:2003.06975, 2020.
[4] O. Adedeji and Z. Wang, “Intelligent waste classification system using
deep learning convolutional neural network,” Procedia Manufacturing,
vol. 35, pp. 607–612, 2019, the 2nd International Conference on
Sustainable Materials Processing and Manufacturing, SMPM 2019, 8-10
Fig. 4. Number of Threads vs Training accuracy March 2019, Sun City, South Africa.
[5] F. A. Azis, H. Suhaimi, and E. Abas, “Waste classification using
convolutional neural network,” in Proceedings of the 2020 International
Conference on Information Technology and Computer Communications,
V. C ONCLUSION AND F UTURE WORK ser. ITCC 2020. Association for Computing Machinery, 2020, p. 9–13.
In this paper, a novel approach incorporating quantum [6] S. Ali and Thung, “Plastic detection model,” https://github.com/
antiplasti/Plastic-Detection-Model, 2022.
learning in transfer learning models, for the purpose of trash [7] A. Masand, S. Chauhan, M. Jangid, R. Kumar, and S. Roy, “Scrapnet:
classification, was presented. A new, comprehensive trash clas- An efficient approach to trash classification,” IEEE Access, vol. 9, pp.
130 947–130 958, 2021.
sification dataset called TrashBox was put together, consisting [8] C. Shi, C. Tan, T. Wang, and L. Wang, “A waste classification
of 17,785 images across various classes. Our dataset is larger method based on a multilayer hybrid convolution neural network,”
than any existing trash classification dataset and also contains Applied Sciences, vol. 11, no. 18, 2021. [Online]. Available:
https://www.mdpi.com/2076-3417/11/18/8572
unique classes such as medical waste and e-waste, hence [9] N. Killoran, T. R. Bromley, J. M. Arrazola, M. Schuld, N. Quesada, and
making the dataset quite diverse in nature. Upon running an S. Lloyd, “Continuous-variable quantum neural networks,” Phys. Rev.
optimized ResNet-101 model, an accuracy of 98.47% was Research, vol. 1, p. 033063, Oct 2019.
[10] A. Mari, T. R. Bromley, J. Izaac, M. Schuld, and N. Killoran, “Transfer
achieved. We hope that TrashBox becomes a valuable resource learning in hybrid classical-quantum neural networks,” Quantum, vol. 4,
to other researchers in their work. The quantum transfer p. 340, Oct 2020.
learning models were applied to the trash classification prob- [11] H. Mogalapalli, M. Abburi, B. Nithya, and S. K. Bandreddi, “Classi-
cal–quantum transfer learning for image classification,” SN Computer
lem, along with optimizations in the form of parallelization Science, vol. 3, no. 1, 2021.
strategies to speed up their training. This improved accuracy [12] R. S. Mor, K. S. Sangwan, S. Singh, A. Singh, and M. Kharub, “E-waste
up to 10.84%, and decreased training time by 27.4%. We management for environmental sustainability: an exploratory study,”
Procedia CIRP, vol. 98, pp. 193–198, 2021, 28th CIRP Conference on
performed several experiments with the designed quantum Lifecycle Engineering, March 10 – 12, 2021, Jaipur, India.
transfer learning models. In particular, several transfer learning [13] M. Tikolo, “Wfdownloader.” [Online]. Available: https://www.
architectures and datasets were experimented with. and best wfdownloader.xyz/
[14] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet:
results were achieved with the Quantum Res-Net-50 model, A large-scale hierarchical image database,” in 2009 IEEE Conference on
which gives 84.97% accuracy with the TrashBox dataset, and Computer Vision and Pattern Recognition, June 2009, pp. 248–255.
80.49% accuracy with the benchmark TrashNet dataset. We [15] S. Rasmussen, N. Loft, T. Bækkegaard, M. Kues, and N. Zinner, “Single-
qubit rotations in parameterized quantum circuits,” arXiv e-prints, pp.
believe that there is scope for further optimizations for this arXiv–2005, 2020.
model, particularly as quantum transfer learning itself is still [16] Xanadu. (2021) Pennylane.ai. [Online]. Available: https://pennylane.ai/

---------------------------------------------------------------------------- 130 ----------------------------------------------------------------------------

Authorized licensed use limited to: Chaitanya Bharathi Institute of Tech - HYDERABAD. Downloaded on August 16,2024 at 03:49:35 UTC from IEEE Xplore. Restrictions apply.