Comparative Study on Binary Waste Classification Based on Deep
Convolutional Neural Networks and Data Augmentation
Shuyuan Xing
a
Information of Technology, The University of New South Wales, Sydney, 1466, Australia
Keywords: Waste Classification, Deep Convolutional Neural Networks, Data Augmentation.
Abstract: With the acceleration of urbanization and the increasing demand for environmental protection, waste
classification has emerged as a crucial component of waste management. This paper proposes three baseline
methods for binary waste classification based on deep convolutional neural networks and data augmentation
techniques. The first baseline employs a pre-trained ResNet50 model combined with an SE attention module
to enhance feature representation; the second baseline utilizes a lightweight EfficientNet-B0 model with
conventional data augmentation strategies; and the third baseline also adopts EfficientNet-B0 but integrates
more aggressive augmentation methods, such as random cropping, color jittering, Gaussian blur, and random
erasing, to improve model generalization. Results from experiments on a Kaggle trash categorization dataset
show that the EfficientNet-B0-based method with aggressive data augmentation significantly increases
accuracy and robustness. This paper serves as a helpful reference for further research in this area since it not
only presents an efficient deep learning solution for waste classification, but it also provides insightful
information about how data augmentation techniques affect model performance.
1 INTRODUCTION
Globally, waste classification has become a crucial
concern for resource recycling and environmental
preservation. The amount of waste produced has
increased due to fast urbanization and
industrialization, making manual sorting techniques
ineffective and prone to mistakes. As a result, they are
unable to satisfy the growing waste management
needs of contemporary cities. In addition to
maximizing resource recovery and lowering pollution
levels, efficient waste classification is crucial for
promoting the circular economy by facilitating the
recycling of valuable materials and lowering
dependency on raw materials. Furthermore, proper
waste classification can promote sustainable
consumption and disposal behaviors by increasing
recycling efficiency, reducing production costs, and
increasing public environmental awareness.
Several technological approaches have been put
forth in recent years to increase the effectiveness of
trash classification. Convolutional Neural Networks
(CNNs), in particular, have shown themselves to be
effective tools for effectively classifying garbage
a
https://orcid.org/0009-0001-7677-6011
photos and extracting rich feature information from
them, greatly increasing identification accuracy
(Ramsurrun et al., 2021). Using cutting-edge
computer vision techniques, recent research has also
investigated two-stage recognition-retrieval strategies
to increase trash categorization accuracy (Zhang, S. et
al., 2021). To further improve classification
performance while reducing the requirement for
sizable, labeled datasets, researchers have also used
CNN optimization techniques and transfer learning
(Zhang, Q. et al., 2021). By combining CNNs with
deep feature refinement techniques, several studies
have investigated hybrid systems and shown
increased classification accuracy for intricate waste
categories (Lu and Chen, 2022). Furthermore, smart
city trash management systems now incorporate
multi-agent simulations and Internet of Things (IoT)
technology, enabling automated and more effective
sorting procedures (Hussain et al., 2024). To improve
real-time waste tracking and maximize resource
allocation, smart waste management solutions also
make use of IoT-based models (Mookkaiah et al.,
2022).
404
Xing, S.
Comparative Study on Binary Waste Classification Based on Deep Convolutional Neural Networks and Data Augmentation.
DOI: 10.5220/0013698300004670
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 2nd International Conference on Data Science and Engineering (ICDSE 2025) , pages 404-408
ISBN: 978-989-758-765-8
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.

This study explores the ways in which
technological advancements can improve trash
classification systems' precision and effectiveness.
Our study intends to improve current approaches and
present innovative ideas to address the problems with
the waste sorting systems that are already in use by
utilizing CNN-based image identification. This will
ultimately help to create a more sustainable waste
management strategy.
2 METHOD
According to recent research, aggressive data
augmentation can greatly increase model robustness
in real-world waste classification settings, especially
when combined with domain adaptation strategies
(Ruiz et al., 2019). Similar results have been found in
earlier research, which showed that deep architectures
like ResNet frequently overfit small datasets,
requiring the use of extra regularization techzniques
such data augmentation or dropout (Ogundana et al.,
2024). Additionally, it has been demonstrated that
using pre-trained deep learning models, like ResNet
and EfficientNet, improves classification
performance while lowering computing costs (Mao et
al., 2021).
2.1 Data Preprocessing and Dataset
Splitting
Three baseline methods were created in this study to
deal with the binary waste classification issue.
Despite having comparable basic functions, the
baselines differed in the model architecture they used
and the degree of data augmentation they used in
preprocessing. The general framework, including data
pretreatment, dataset partitioning, model creation,
and training and evaluation techniques, is described
in detail in the parts that follow.
Image data were first preprocessed using a
number of common transformations for each of the
three baselines. All photos in Baselines 1 and 2 were
resized to 224 x 224 pixels as part of the
preprocessing pipeline. This was followed by random
rotation (up to 15°), random horizontal flipping, and
color jittering (changing contrast and brightness) to
add unpredictability. After that, images were
transformed into tensors and normalized using the
ImageNet dataset's mean and standard deviation. In
order to further improve the classifier's resilience, a
more aggressive augmentation method was used for
Baseline 3. In addition to the usual random horizontal
flip and rotation, this method included a randomly
scaled crop to 224×224 with customizable scale and
aspect ratio. Additionally, Baseline 3 used random
erasing to mimic occlusions, Gaussian blur with a
changeable sigma, and more noticeable color jittering
(including saturation and hue adjustments).
2.2 Model Architectures
Throughout the baselines, two main deep learning
architectures were used. A ResNet50 model that has
already been trained on ImageNet is used in Baseline
1. In this configuration, the original final fully
connected layer of ResNet50 is replaced by a new
linear layer with a single output neuron to
accommodate binary classification. An optional
Squeeze-and-Excitation (SE) module is integrated
into the network to recalibrate channel-wise feature
responses and enhance the discriminative capacity of
the extracted features.
In contrast, Baseline 2 and Baseline 3 adopt the
EfficientNet-B0 architecture, which is known for its
efficiency and scalability. Similar to the ResNet50
modification, the final classifier layer of EfficientNet-
B0 is replaced with a linear layer producing a single
output. Although both baselines use EfficientNet-B0
as the backbone, Baseline 2 applies the standard data
augmentation strategy (similar to Baseline 1), while
Baseline 3 incorporates the enhanced, aggressive
augmentation pipeline described above. In both
EfficientNet-based models, an optional SE module
can be included, though the default configuration did
not enable attention in our experiments.
2.3 Training and Evaluation
Procedures
The Adam optimizer was used for training, with an
initial learning rate of 1×10⁻⁴ for each baseline. The
loss function was defined as Binary Cross Entropy
with Logits (BCEWithLogitsLoss), which is
appropriate for binary classification tasks since it
naturally blends a Sigmoid activation with cross-
entropy loss. The model's performance on the training
set was tracked by computing the average loss and
accuracy over the 50 epochs that the training process
was iterated over. After each epoch, the model was
evaluated on the validation set to assess
generalization performance.
A progress bar that showed real-time loss numbers
was used to monitor training progress. Furthermore, a
logging mechanism was put in place to capture the
model architecture specifics as well as the accuracy
and loss metrics at each epoch to a log file. To give an
Comparative Study on Binary Waste Classification Based on Deep Convolutional Neural Networks and Data Augmentation
405

objective assessment of the model's performance, it
was further tested on a different test set after training.
The training and validation metrics were
subsequently visualized using line plots to illustrate
convergence behavior and model stability.
3 RESULTS
3.1 Baseline 1: ResNet50 + SE
Rapid convergence on the training set is indicated by
the Baseline 1 (ResNet50 with the SE attention
module) training and validation curves, which are
displayed in figures 1 and 2. An outstanding fit to the
training data is shown by the training accuracy
reaching and remaining above 99% and the training
loss rapidly decreasing to almost nil.
In contrast, the validation loss fluctuates between
approximately 0.15 and 0.25, with the validation
accuracy stabilizing around 93%–95%. These
observations suggest that while the model is highly
effective on the training set, there is some degree of
overfitting. The final test results for Baseline 1
are Test Loss = 0.2858 and Test Accuracy = 93.63%,
indicating that the ResNet50 + SE configuration is
capable of robust feature extraction and generalizes
well to unseen data despite the slight overfitting.
Figure 1: ResNet50 + SE Loss Curve. (Picture credit:
Orignial)
Figure 2: ResNet50 + SE Accuracy Curve. (Picture credit:
Orignial)
3.2 Baseline 2: EfficientNet-B0 +
Standard Data Augmentation
Baseline 2 employs the lightweight EfficientNet-B0
model along with a standard data augmentation
strategy similar to that of Baseline 1. The training
curves for Baseline 2 show rapid convergence, with
the training loss quickly approaching zero and the
training accuracy nearing 100%. However, the
validation loss remains consistently higher,
oscillating between 0.13 and 0.22, while the
validation accuracy stays in the range of 94%–95%
shown as figure 3 and 4. The final test results for
Baseline 2 are Test Loss = 0.4420 and Test Accuracy
= 93.16%. These metrics suggest that, although
EfficientNet-B0 provides a faster and more
lightweight alternative, its ability to generalize to the
test set is slightly compromised when only
conventional data augmentation is applied.
Figure 3: EfficientNet-B0 + Standard Data Augmentation
Loss Curve. (Picture credit: Orignial)
ICDSE 2025 - The International Conference on Data Science and Engineering
406

Figure 4: EfficientNet-B0 + Standard Data Augmentation
Accuracy Curve. (Picture credit: Orignial)
3.3 Baseline 3: EfficientNet-B0 +
Aggressive Data Augmentation
In Baseline 3, the same EfficientNet-B0 backbone is
used; however, a more aggressive data augmentation
pipeline is employed. This pipeline includes random
resized cropping, stronger color jitter (with
adjustments in saturation and hue), Gaussian blur, and
random erasing. The training curves again show rapid
convergence with the training loss approaching zero
and near-perfect training accuracy. The validation
loss for Baseline 3 remains within a similar range as
in the previous baselines (approximately 0.15–0.25),
and the validation accuracy is observed to be around
96% shown as figure 5 and 6. The final test results
are Test Loss = 0.2964 and Test Accuracy = 93.51%,
which are very close to those of Baseline 1 and
slightly better than Baseline 2. These results
demonstrate that the incorporation of aggressive data
augmentation significantly enhances the
generalization ability of the lightweight EfficientNet-
B0 model, bringing its performance close to that of
the deeper ResNet50 + SE model.
Figure 5: EfficientNet-B0 + Aggressive Data Augmentation
Loss Curve. (Picture credit: Orignial)
Figure 6: EfficientNet-B0 + Aggressive Data Augmentation
Accuracy Curve. (Picture credit: Orignial)
3.4 Result Analysis
The paper designed and implemented three baseline
methods for binary waste classification, employing a
pre-trained ResNet50 with an SE attention module, an
EfficientNet-B0 model with standard data
augmentation, and an EfficientNet-B0 model
enhanced with aggressive data augmentation.
Experimental results demonstrated that all models
converged rapidly during training, achieving nearly
100% training accuracy; on the test and validation sets,
overfitting was noted, though. With a test accuracy of
93.63% and a test loss of 0.2858, the ResNet50 + SE
model in particular demonstrated that deep networks
and attention mechanisms may work together to
efficiently extract picture data and perform well in
waste classification tasks. In contrast, the
EfficientNet-B0 model with standard data
augmentation reached only 93.16% accuracy and a
higher test loss of 0.4420, reflecting slightly weaker
generalization performance. Significantly, the
EfficientNet-B0 model's ability to adapt to a variety
of image data was enhanced by implementing
aggressive data augmentation strategies. It achieved a
test accuracy of 93.51% and a significantly lower test
loss of 0.2964, almost matching the ResNet50 + SE
model's performance while providing advantages in
terms of model size and computational efficiency, as
indicated in table 1.
Comparative Study on Binary Waste Classification Based on Deep Convolutional Neural Networks and Data Augmentation
407

Table 1: Test result of three baseline model.
Loss Accurac
y
ResNet50 + SE 0.2858 93.63%
EfficientNet-
B0 + Standard
Data
Augmentation
0.4420 93.16%
EfficientNet-
B0 +
Aggressive
Data
Au
g
mentation
0.2964 93.51%
4 CONCLUSIONS
Overall, our study demonstrates that while deep
architectures like ResNet50 + SE are capable of
extracting rich image features and achieving high
classification accuracy, lightweight models such as
EfficientNet-B0 can attain comparable performance
when enhanced with advanced data augmentation
techniques, making them particularly suitable for
resource-constrained applications. Data augmentation
is shown to play a critical role in improving model
generalization by effectively mitigating overfitting
and ensuring robust performance on unseen data.
In order to further minimize overfitting, future
research will investigate more complex regularization
techniques and dynamic learning rate scheduling. To
reduce overfitting in trash categorization models,
several optimization strategies have been put forth,
including adaptive weight decay and dynamic
learning rate scheduling. Furthermore, it is
anticipated that adding attention mechanisms or other
sophisticated feature reconstruction methods to
lightweight models may improve classification
performance even further without requiring a large
amount of processing cost. To improve the
robustness and generalization of waste classification
systems in complicated circumstances, efforts will
also be made to expand and refine the dataset, explore
semi-supervised or unsupervised learning
methodologies, and use cross-domain data fusion
techniques. In the end, these study developments will
offer technical assistance for the implementation of
automated waste sorting systems that is more
effective, economical, and long-lasting.
REFERENCES
Hussain, I., Elomri, A., Kerbache, L., and El Omri, A. 2024.
Smart city solutions: Comparative analysis of waste
management models in IoT-enabled environments
using multiagent simulation. Sustainable Cities and
Society, 103, 105247. doi: 10.1016/j.scs.2024.105247.
Jin, S., Yang, Z., Królczyk, G., Liu, X., Gardoni, P., and Li,
Z. 2023. Garbage detection and classification using a
new deep learning-based machine vision system as a
tool for sustainable waste recycling. Waste
Management, 162, 123–130. doi:
10.1016/j.wasman.2023.02.014.
Lu, W., and Chen, J. 2022. Computer vision for solid waste
sorting: A critical review of academic research. Waste
Management, 142, 29–43. doi:
10.1016/j.wasman.2022.02.009.
Mao, W.-L., Chen, W.-C., Wang, C.-T., and Lin, Y.-H.
2021. Recycling waste classification using optimized
convolutional neural network." Resources,
Conservation and Recycling, 164, 105132. doi:
10.1016/j.resconrec.2020.105132.
Mookkaiah, S. S., Thangavelu, G., Hebbar, R., et al. 2022.
"Design and development of smart Internet of Things–
based solid waste management system using computer
vision." Environmental Science and Pollution Research,
29, 64871–64885. doi: 10.1007/s11356-022-20428-
2.Goals (SEB4SDG), Omu-Aran, Nigeria, 1–11. doi:
10.1109/SEB4SDG60871.2024.10629933.
Ogundana, A. K., Afolabi, O. O., Ilevbare, M., and Falae, P.
O. (2024). "Green Hydrogen Generation from Plastic
Waste: A Review." In 2024 International Conference on
Science, Engineering and Business for Driving
Sustainable Development Goals (SEB4SDG), Omu-
Aran, Nigeria, 1–11. doi:
10.1109/SEB4SDG60871.2024.10629933.
Ramsurrun, N., Suddul, G., Armoogum, S. and Foogooa, R.
2021. Recyclable Waste Classification Using Computer
Vision And Deep Learning. in 2021 Zooming
Innovation in Consumer Technologies Conference
(ZINC), Novi Sad, Serbia, pp. 11–15, doi:
10.1109/ZINC52049.2021.9499291.
Ruiz, V., Sánchez, Á., Vélez, J. F., and Raducanu, B. 2019.
"Automatic Image-Based Waste Classification." In
From Bioinspired Systems and Biomedical
Applications to Machine Learning. IWINAC 2019,
Lecture Notes in Computer Science, 11487, edited by
Ferrández Vicente, J., Álvarez-Sánchez, J., de la Paz
López, F., Toledo Moreo, J., and Adeli, H. Springer,
Cham. doi: 10.1007/978-3-030-19651-6_41.
Zhang, Q., Yang, Q., Zhang, X., Bao, Q., Su, J., and Liu, X.
2021. Waste image classification based on transfer
learning and convolutional neural network. Waste
Management, 135, 150–157. doi:
10.1016/j.wasman.2021.08.038.
Zhang, S., Chen, Y., Yang, Z., and Gong, H. 2021.
Computer Vision Based Two-stage Waste Recognition-
Retrieval Algorithm for Waste Classification.
Resources, Conservation and Recycling, 169, 105543.
doi: 10.1016/j.resconrec.2021.105543.
ICDSE 2025 - The International Conference on Data Science and Engineering
408