Multi-label Image Classification using Deep Learning

Algorithms
Muhammed Al Rashid (12007142)
Department of Computer Science and Engineering,
Lovely professional university.
Delhi-Jalandhar GT Road, Phagwara, Punjab, India(144001)
[email protected]
Dr. Soni Singh
Department of Computer Science and Engineering,
Lovely professional university.
Delhi-Jalandhar GT Road, Phagwara, Punjab, India(144001)

Abstract - This research paper explores the application of base by conducting a systematic evaluation and comparison of
deep learning techniques for multi-label image classification, several models in order to determine the best appropriate
utilizing three distinct models: Multi-Layer Perceptron (MLP), technique for multi-label picture classification.
Convolutional Neural Network (CNN), and MobileNetV2 with
Transfer Learning. The study investigates the performance of
these models on a multi-label image classification task, aiming to
One of the fundamental aspects of our study involves
achieve superior accuracy and comparative analysis. Two conducting a comparative examination of optimisation
optimization algorithms, Adam and Stochastic Gradient Descent approaches, specifically examining the effects of two widely
(SGD), are employed to enhance model training and evaluate used algorithms: Adam and Stochastic Gradient Descent
their impact on classification performance. The paper presents a (SGD). The optimisation of deep learning models is of utmost
comprehensive analysis of the experimental results, highlighting importance in order to attain both high accuracy and expedited
the strengths and weaknesses of each model and optimization convergence throughout the training process. The selection of
method, providing valuable insights for researchers and an optimisation technique can significantly impact the
practitioners in the field of deep learning-based image performance of a model. In this study, we seek to provide
classification.
insights into the advantages and limitations of both Adam and
Index Terms – Transfer learning, CNNs, Computer Vision,
SGD algorithms within the domain of multi-label picture
Deep Learning, Multi-label Image Classification classification.

1. INTRODUCTION The models and techniques chosen for this study encompass a
wide array of methodologies, ranging from a basic Multilayer
In an era characterized by the prevalence of visual Perceptron (MLP) to an advanced Convolutional Neural
content, the capacity to automatically identify and categorise Network (CNN), and incorporating the utilization of pre-
things depicted in photographs has emerged as a significant trained models via transfer learning. The presence of diversity
obstacle in the fields of computer vision and artificial within the dataset provides an opportunity to evaluate the
intelligence. The task of multi-label image classification is of compromises that arise from factors such as the intricacy of
significant importance in various applications such as content the model, the computing demands, and the effectiveness of
recommendation, medical diagnosis, and object recognition. It the classification. The primary aim of this study is to offer a
involves the assignment of multiple labels or tags to an image. comprehensive comprehension of the multi-label image
The utilisation of deep learning has arisen as a potent classification field to researchers, practitioners, and machine
mechanism for tackling this particular difficulty, as it learning enthusiasts. This will enable them to make well-
possesses the ability to acquire intricate features from data and informed choices regarding the selection of models and
substantially enhance the accuracy of categorization. This optimization techniques that are most suitable for their
research paper aims to investigate the field of multi-label particular applications.
picture classification and assess the effectiveness of three
well-known deep learning models: Multi-Layer Perceptron In this study, we show comprehensive experimental findings,
(MLP), Convolutional Neural Network (CNN), and Transfer analyse the intricacies of model training, and offer valuable
Learning with pre-trained MobileNetV2. observations regarding the optimization process of Adam and
SGD. The implications of this research are expected to
The rationale behind conducting this research arises from the contribute significantly to the progress of multi-label picture
increasing demand to improve the accuracy of multi-label classification and provide a valuable reference for individuals
picture categorization systems. As the proliferation and aiming to leverage deep learning in many practical contexts.
intricacy of visual data persistently expand, conventional Our objective is to provide a contribution to the current
approaches frequently prove inadequate in delivering precise endeavours in enhancing the accessibility and efficacy of deep
and expedient solutions. Deep learning models have exhibited learning as a tool for picture classification problems, through
exceptional abilities in addressing picture classification the integration of theoretical knowledge with practical
challenges; nonetheless, the process of choosing a suitable experiments..
model and optimization technique is not straightforward. The
objective of this study is to enhance the current knowledge
 framework by the authors introduces a semantic-
2. LITERATURE REVIEW aware dual contrastive learning approach that
integrates two distinct types of contrastive learning:
Numerous researchers worldwide have delved into the sample-to-sample contrastive learning (SSCL) and
realm of multi-label image classification, contributing to a rich prototype-to-sample contrastive learning (PSCL).
body of scholarly work. In the following sections, we provide The proposed methodology involves the extraction of
concise reviews and insights into a selection of research local discriminative characteristics that are relevant
papers on this subject. to specific categories. These features are then used to
create category prototypes. Additionally, the strategy
1. Comparative Study Of Deep Learning Models In aggregates visual representations at the label level
Multi-label Scene Classification: [1] The and effectively differentiates features belonging to
investigation of computer vision applications for different categories. The framework additionally
environmental monitoring, namely the utilisation of establishes a PSCL module with the purpose of
machine learning techniques on remotely sensed reducing the disparity between positive samples and
images, is currently under examination. The dataset category prototypes. The efficacy of the suggested
utilised in this study is the UC Merced Land Use strategy is demonstrated by experiments conducted
dataset, which was employed for the purpose of on five extensive public datasets.
conducting a comparative analysis of seven deep
learning models in the context of multi-class 5. Feature learning network with transformer for
classification. The DenseNet 121 model multi-label image classification: [5] The authors
demonstrated the highest level of accuracy, although provide a novel approach called FL-Tran, which is a
the Alexnet and SqueezeNet models exhibited Feature Learning network based on the Transformer
superior predictive capabilities. In subsequent architecture. The purpose of this approach is to
research endeavours, the utilisation of domain-shift enhance the performance of multi-label image
apps could be employed to explore broader domains. classification. The proposed network integrates many
modules to boost its performance. Firstly, a multi-
2. Residual Attention: A Simple but Effective scale fusion module (MSFM) is employed to align
Method for Multi-Label Recognition: [2] This high-level and low-level features. Secondly, a spatial
study introduces the concept of class-specific residual attention module (SAM) is utilised to capture
attention (CSRA) as a straightforward module for the prominent object characteristics. Lastly, a feature
purpose of multi-label image recognition. CSRA enhancement and suppression module (FESM) is
employs a methodology to generate features that are incorporated to extract hidden valuable information.
distinct to each class, resulting in superior The performance of the FL-Tran model in learning
performance compared to existing approaches. This and identifying small-scale objects has been
approach demonstrates consistent enhancements demonstrated by experiments conducted on the MS-
across a wide range of models and datasets. The COCO 2014, PASCAL VOC 2007, and NUS-WIDE
implementation of this solution is straightforward, as datasets. These experiments have shown that the FL-
it is designed to be lightweight and offers user- Tran model surpasses existing approaches in this
friendly explanations and visual representations. domain.

3. Deep Convolution Neural Network sharing for the 6. Boosting Multi-Label Image Classification with
multi-label images classification: [3] This study Complementary Parallel Self-Distillation: [6] The
introduces a multi-label classification framework that research paper introduces a theoretical construct
utilises the Multi-Branch Neural Network Model known as Parallel Self-Distillation (PSD) with the
(MBNN) to encode input from multiple semi-parallel aim of enhancing the performance of Multi-Label
subnetworks or layers outputs individually. The Image Classification (MLIC) models. This is
architectural design incorporates subnetworks based achieved through the decomposition of complex tasks
on Convolutional Neural Networks and may be into more manageable sub-tasks, the acquisition of
flexibly adjusted to accommodate Multitask Learning joint and category-specific patterns, and the
topologies. The suggested architecture demonstrates utilisation of knowledge distillation to effectively
superior performance compared to alternative basic balance the exploitation of label correlations and the
multi-label classification methods, with the "network prevention of model overfitting.
with multi-features" getting the highest classification
score. 7. Visual Transformers with Primal Object Queries
for Multi-Label Image Classification: [7] This
4. Semantic-Aware Dual Contrastive Learning for research suggests the incorporation of primal object
Multi-label Image Classification: [4] The proposed queries into a vision-based transformer model to
 enhance the performance of multi-label image
classification. Object queries, which are positional 3. MULTI-LAYER PERCEPTRON
encodings that may be learned, are utilised for the
purpose of decoding object classes or bounding The Multi-Layer Perceptron (MLP) is a widely employed
boxes. The proposed model demonstrates artificial neural network in the field of machine learning,
enhancements in the F1 measure, surpassing the particularly in problems such as multi-label image
current state-of-the-art by 2.1% and 1.8% on the MS- classification. It is recognised for its simplicity and
COCO and NUS-WIDE datasets, respectively. effectiveness in these applications. The neural network
Additionally, the model exhibits accelerated architecture comprises an input layer, one or more hidden
convergence, reducing the required training time by layers, and an output layer, whereby neurons are
79.0% and 38.6% on the aforementioned datasets. interconnected by weighted connections.

8. Benchmarking and scaling of deep learning The MLP with kernel regularisation, in its expanded form,
models for land cover image classification: [8] The employs various regularisation strategies to enhance
utilisation of Copernicus Sentinel-2 imagery has generalisation and alleviate the issue of overfitting.
novel prospects for the application of deep learning Techniques like as L1 or L2 regularisation incorporate an
techniques in the classification of land use and land additional term in the loss function, which serves to promote
cover images. This study evaluates the performance simpler weight configurations and mitigate excessive attention
of 60 contemporary deep learning models using the on certain neurons or features. This approach aids in managing
BigEarthNet Sentinel-2 dataset. The models the complexity of the network and mitigating the risk of
encompass several types such as standard overfitting by deterring the network from excessively fitting
convolutional neural networks (CNNs), noise present in the training data.
EfficientNets, and Wide Residual Networks. The
utilisation of the model zoo facilitates the application The incorporation of kernel regularisation has been observed
of transfer learning and expedites the process of to enhance performance, particularly in intricate multi-label
prototyping in activities related to remote sensing. image classification problems and datasets with restricted
availability. The selection of the regularisation strength
9. A Study on CNN Transfer Learning for Image parameter holds significant importance as it governs the
Classification: [9] The present study aims to balance between effectively fitting the training data and
examine the efficacy of the Convolutional Neural mitigating the risk of overfitting.
Network (CNN) architecture model known as
Inception-v3 in the task of picture classification,
specifically through the utilisation of Transfer
Learning. The model undergoes evaluation to assess
its accuracy and efficiency using novel image
datasets, and its performance is compared to that of
cutting-edge methodologies in the field of Computer
Vision.

10. Multilabel Image Classification with Deep

Transfer Learning for Decision Support on
Wildfire Response: [10] The present study aimed to
devise a multilabel classification (MLC) model that
Figure 1. Multi-layer perceptron (MLP-NN) Architecture [11].
incorporates transfer learning and data augmentation
techniques in order to overcome the constraints
4. CONVOLUTIONAL NEURAL NETWORK
observed in current algorithms for forest fire
detection. The model generates many pieces of
The Convolutional Neural Network (CNN) is a type of deep
information pertaining to a given item or image,
learning framework that has been specifically developed for
hence mitigating the potential issue of overfitting
computer vision tasks, with a particular focus on image
caused by limited datasets. The proposed model has
classification. In the context of multi-label image
the potential to provide valuable guidance for future
classification, convolutional layers are employed to
research endeavours focused on the implementation
automatically extract significant characteristics from the
of deep learning-based field image analysis and
images.
decision support systems within the context of
wildfire response efforts.
The heart of CNN architecture consists of several
convolutional layers that employ convolution operations to
identify patterns and features at various sizes. These layers
employ stacked learnable filters or kernels to facilitate related to picture categorization. MobileNetV2 employs depth
intricate and hierarchical learning processes. wise separable convolutions and many architectural
optimizations to attain a harmonious trade-off between
Dropout layers are a form of regularization employed in accuracy and efficiency.
Convolutional Neural Networks (CNNs) to introduce
randomness by deactivating a proportion of neurons during the When implementing transfer learning with MobileNetV2, it is
training process. This technique serves the purpose of customary to adhere to the subsequent procedures:
mitigating overfitting and promoting the development of
resilient features. These methods seem to be beneficial in the Pre-trained MobileNetV2: The MobileNetV2 model that has
context of extensive networks or when confronted with a been pre-trained. MobileNetV2 has undergone pre-training
scarcity of training data. using a comprehensive image dataset, such as ImageNet,
which encompasses a substantial number of images. During
Kernel regularization is an essential element in convolutional the initial training phase, the model acquires the ability to
neural network (CNN) models. It entails the inclusion of a extract comprehensive and universal features from a diverse
penalty term in the loss function during the training process. range of images.
This penalty term serves to encourage the development of
smaller and more uniform weights, so effectively managing Extraction of Features: In order to employ MobileNetV2 for a
the issue of overfitting. Two often used methods of novel picture classification undertaking, one can leverage the
regularization are L1 and L2 regularization. insights encapsulated inside the pre-trained layers of the
model. The feature extraction process involves the removal of
This study examines the efficacy of Convolutional Neural the uppermost classification layers of MobileNetV2, hence
Network (CNN) models that use Dropout layers and kernel enabling the utilization of these layers for this purpose.
regularization techniques in the context of multi-label picture
classification tasks. The inclusion of dropout layers in a model Transfer Learning: Transfer learning is a technique in machine
enhances its ability to resist overfitting, particularly when learning where knowledge gained from one task is used to
dealing with intricate datasets. On the other hand, taking another related task. It involves using pre-trained models as a
advantage of kernel regularization techniques guarantees that starting point and then fine Subsequently, more layers are
the learned features are not excessively intricate, hence incorporated into the pre-existing MobileNetV2 model in
improving the model's generalization capabilities. The order to customize it for the particular task at hand. The
objective of this study is to achieve a balance between model further layers encompass the output layer, which aligns with
complexity and generalization, leading to enhanced accuracy the total number of classes present within the dataset.
in the classification of images with multiple labels.. Typically, the initial layer of MobileNetV2 is initialized in a
random manner, while the remaining layers are fixed, so
ensuring that they remain unaltered during the training
process.

Fine-Tuning: Fine-tuning is a process in which a pre-trained

model is further trained on a specific task or dataset to
Figure 2 Dropout in Regu CNN [12] improve its performance. In certain instances, it is possible to
employ fine-tuning on the pre-trained MobileNetV2 layers by
5. TRANSFER LEARNING USING MOBILENET utilizing a reduced learning rate. The process of fine-tuning
enables the model to adjust to the individual attributes of the
Transfer learning is a prominent machine learning dataset, while still keeping the useful information acquired
methodology wherein a pre-existing model is employed as an from the pre-trained layers.
initial framework for a novel task or dataset, as opposed to
training a model from its inception. This methodology utilizes Transfer learning using the MobileNetV2 model presents
the knowledge and feature representations that a pre-trained numerous benefits:
model has learned from a substantial and varied dataset.
MobileNetV2, a widely utilized pre-trained model, holds Efficiency: The concept of efficiency refers to the ability to
significant popularity in the domain of picture classification, accomplish a task or achieve a goal with minimal waste of
particularly in the application of transfer learning. resources, such as time, MobileNetV2 exhibits high efficiency
in terms of its model size and inference performance,
MobileNetV2 is a convolutional neural network architecture rendering it well-suited for deployment in contexts with
specifically developed for the purpose of mobile and limited resources.
embedded vision applications. The model is recognized for its
efficacy in terms of its compact size and processing demands, Enhanced Performance: By utilizing the pre-trained
while simultaneously exhibiting robust performance in tasks MobileNetV2 features, it is possible to attain superior
performance and expedite convergence for the new picture
classification task, even when confronted with a scarcity of 7. METHODOLOGY
data.
7.1 Load and Preprocessing Image:
Reduced Training Time: One of the advantages of this
approach is the reduction in training time. Given the initial  To facilitate the training process, it is important to
robust feature set, the training procedure necessitates a obtain a dataset initially. The animal dataset was
reduced number of epochs in comparison to training the model obtained via Kaggle. I incorporated the Kaggle API
from its inception. into a Google Colab notebook.

This methodology is extensively employed in many  Next, the dataset was populated with over 9000
applications, including object recognition, fine-grained picture photos, which were used for both training and testing
classification, and other related domains. This approach purposes. Subsequently, the photos were subjected to
enables users to leverage the extensive knowledge embedded preprocessing and scaled to dimensions of 160x160
within pre-trained models such as MobileNetV2, while pixels, as the original photographs exhibited varying
simultaneously customizing the model to suit their individual sizes.
requirements.
 Additionally, the image was rescaled to a range of 0.0
6. OPTIMIZATION ALGORITHM to 1.0. Subsequently, various transformations such as
rotation, zoom, shift, and shear were applied.
Adam (short for Adaptive Moment Estimation) and Stochastic Furthermore, the dataset was separated into a training
Gradient Descent (SGD) are two popular optimization set and a validation set in a ratio of 0.8:0.2,
algorithms used in training machine learning models, respectively, for the purpose of training.
including deep learning neural networks. Each has its own
characteristics and advantages, making them suitable for  There are five classes such as 'elefante', 'farfalla',
different scenarios 'mucca', 'pecora', 'scoiattolo' in the dataset

6.1 ADAM: Adam is renowned for its capacity to dynamically

adjust the learning rate throughout the training process. The
algorithm calculates distinct learning rates for each parameter
within the model. The capacity to adapt is advantageous in
scenarios where certain parameters require more frequent
updates than others, resulting in expedited convergence.

Adam also includes a momentum term, which bears

resemblance to the concept of stochastic gradient descent
(SGD) with momentum. This facilitates the optimisation
process by enabling the model to traverse smoother regions
and efficiently evade local minima.

6.2 Stochastic Gradient Descent (SGD): The stochastic

gradient descent (SGD) algorithm is well recognised as the
conventional optimisation method employed in the training of
neural networks. The model parameters are updated by
utilising the gradients of the loss function with respect to the
parameters.

In contrast to Adam, the Stochastic Gradient Descent (SGD)

algorithm necessitates meticulous adjustment of the learning
rate. The selection of an adequate learning rate is of utmost
importance in ensuring the stability and convergence of 7.2 Train using Multi-Layer Perceptron Model:
training. Excessively high learning rates have the potential to
cause divergence, whilst excessively low learning rates may  A multilayer perceptron (MLP) model was
lead to sluggish convergence. constructed, incorporating kernel regularisation.
Activation functions such as rectified linear unit
(ReLU) and softmax were employed for the dense
and dropout layers.
  The optimisation process employed the Stochastic
Gradient Descent (SGD) algorithm in conjunction
with a learning rate scheduler, which systematically
reduced the learning rate at predetermined intervals.

 Additionally, EarlyStopping was used to cease model

training when no substantial improvement isobserved,
allowing for progression to subsequent steps.

 We received the accuracy of 0.4938 and validation

accuracy of 0.5366, which is very lower for the
dataset.

7.4. Prediction and Transfer Learning:

 In this study, the pre-trained MobileNetV2 model was

utilised as the foundational model for prediction
purposes. Next, an Input layer was incorporated into
the model with the proper shape.

 After the basic model, a Global Average Pooling 2D

layer is applied, which is then followed by a Dense
layer serving as the final prediction layer. The model
demonstrated strong performance with an accuracy of
0.91 and a validation accuracy of 0.9 after 10 epochs.
In this study, the ADAM optimizer was employed
once again, demonstrating superior performance
compared to both the CNN and MLP models..

∞
���
7.3 Train using CNN Model:� � = �0 + �� cos
�
+
�=1
���
�� sin
�

 In the subsequent phase, we conducted training

utilizing a Convolutional Neural Network (CNN)
model that incorporated a basic Dropout layer and
Kernel Regularization. In order to optimize the
model, the Adam optimizer is employed, along with
the implementation of EarlyStopping and
ReduceLearningRateOnPlateau techniques.

 The accuracy of the CNN model is 0.6, and the

validation accuracy is also 0.6. In both cases, it is
evident that the CNN model exhibits superior
performance compared to the simple MLP model.
 7. RESULT In summary, the findings of this study emphasize the
In this study, we investigated the performance of three following key takeaways:
distinct models for multi-label image classification: the Multi-
Layer Perceptron (MLP), the Convolutional Neural Network 1. The choice of model architecture is pivotal. More complex
(CNN), and MobileNetV2 with transfer learning. All models models like CNNs and pre-trained models with transfer
were trained for 10 epochs using the Adam optimizer. The learning, such as MobileNetV2, are well-suited for multi-label
results reveal significant differences in accuracy and highlight image classification tasks due to their ability to learn intricate
the effectiveness of transfer learning with MobileNetV2 in features and patterns.
enhancing multi-label image classification.
2. The selection of the optimizer plays a significant role in
1) MLP Model with Stochastic Gradient Descent (SGD): model training. The Adam optimizer's adaptive learning rates
- The initial experiment using the Multi-Layer Perceptron proved highly effective in enhancing model convergence and
(MLP) model with Stochastic Gradient Descent (SGD) as the performance.
optimizer yielded an accuracy of 0.3. This relatively low
accuracy can be attributed to the limitations of a simple MLP 3. Transfer learning, as exemplified by MobileNetV2, is a
architecture when dealing with complex image classification powerful technique for leveraging pre-trained models and
tasks. The model struggled to learn the intricate features significantly improving accuracy in multi-label image
necessary for multi-label image classification. classification tasks.

2) CNN Model with Adam Optimizer: Overall, these insights can guide researchers and practitioners
- In the subsequent experiment, we transitioned to a more in selecting the most suitable approach for their specific image
complex CNN model and employed the Adam optimizer. This classification projects, ultimately leading to more accurate and
change resulted in a significant improvement in accuracy, reliable results.
achieving a score of 0.6. The CNN model's ability to learn 8. REFERENCES
hierarchical features and the adaptability of the Adam [1] Atik, Saziye. (2022). COMPARATIVE STUDY OF DEEP LEARNING
optimizer played a crucial role in boosting the classification MODELS IN MULTI-LABEL SCENE CLASSIFICATION.
performance. [2] Zhu, K., & Wu, J. (2021, August 5). Residual Attention: A Simple but
Effective Method for Multi-Label Recognition. arXiv.org.
https://arxiv.org/abs/2108.02456v2
3) MobileNetV2 with Transfer Learning and Adam Optimizer: [3] Deep Convolution Neural Network sharing for the multi-label images
- The final and most notable experiment involved training classification. (2022, October 26). Deep Convolution Neural Network
the MobileNetV2 architecture with transfer learning using the Sharing for the Multi-label Images Classification - ScienceDirect.
Adam optimizer. This approach yielded the highest accuracy https://doi.org/10.1016/j.mlwa.2022.100422
[4] Ma, L., Sun, D., Wang, L., Zhao, H., & Luo, B. (2023, July 19).
of 0.9. Transfer learning with MobileNetV2 allowed us to Semantic-Aware Dual Contrastive Learning for Multi-label Image
leverage the pre-trained model's extensive knowledge, Classification. arXiv.org. https://arxiv.org/abs/2307.09715v4
enhancing the model's ability to capture complex features and [5] Feature learning network with transformer for multi-label image
patterns within the image dataset. This significantly improved classification. (2022, November 23). Feature Learning Network With
Transformer for Multi-label Image Classification - ScienceDirect.
the accuracy and demonstrated the power of transfer learning https://doi.org/10.1016/j.patcog.2022.109203
in multi-label image classification. [6] Xu, J., Huang, S., Zhou, F., Huangfu, L., Zeng, D., & Liu, B. (2022, May
23). Boosting Multi-Label Image Classification with Complementary
8. CONCLUSION Parallel Self-Distillation. arXiv.org. https://arxiv.org/abs/2205.10986v1
This research underscores the critical importance of [7] Yazici, V. O., de Weijer, J. V., & Yu, L. (2021, December 10). Visual
Transformers with Primal Object Queries for Multi-Label Image
choosing the right model and optimization technique for Classification. arXiv.org. https://arxiv.org/abs/2112.05485v2
multi-label image classification tasks. The results clearly [8] Papoutsis, I., Bountos, N. I., Zavras, A., Michail, D., & Tryfonopoulos, C.
illustrate the limitations of a simple MLP model, which (2021, November 18). Benchmarking and scaling of deep learning models
struggled to achieve satisfactory accuracy on the given dataset. for land cover image classification. arXiv.org.
https://arxiv.org/abs/2111.09451v3.
[9] Hussain, M., Bird, J. J., & Faria, D. R. (2018, August 11). A Study on
Transitioning to a CNN model with the Adam optimizer CNN Transfer Learning for Image Classification. A Study on CNN
significantly improved the classification accuracy, Transfer Learning for Image Classification | SpringerLink.
highlighting the capacity of deep learning architectures in https://doi.org/10.1007/978-3-319-97982-3_16
[10]Park, M., Tran, D. Q., Lee, S., & Park, S. (2021, October 5). Multilabel
image classification tasks. However, the most striking Image Classification with Deep Transfer Learning for Decision Support
outcome was achieved when employing MobileNetV2 with on Wildfire Response. MDPI. https://doi.org/10.3390/rs13193985
transfer learning, also using the Adam optimizer. This [11]Afan, Haitham & Ibrahem Ahmed Osman, Ahmedbahaaaldin & Essam,
approach not only outperformed the other models but achieved Yusuf & Najah, Al-Mahfoodh & Huang, Yuk & Kisi, Ozgur & Sherif,
Mohsen & Sefelnasr, Ahmed & Chau, Kwok & El-Shafie, Ahmed.
an accuracy level of 0.9, demonstrating the transformative (2021). Modeling the fluctuations of groundwater level by employing
impact of transfer learning in this context. ensemble deep learning techniques. Engineering Applications of
Computational Fluid Mechanics. 15. 1420-1439.
10.1080/19942060.2021.1974093.
[12]Hou, Saihui and Zilei Wang. “Weighted Channel Dropout for
Regularization of Deep Convolutional Neural Network.” AAAI
Conference on Artificial Intelligence (2019)..