Comparative Analysis of VGG16 and EfficientNet for Image-Based

Cat Breed Classification

Hao Xu

Polytechnic Institute, Purdue University, West Lafayette, Indiana, 47907, U.S.A.

Keywords: Cat Bread Classification, Convolutional Neural Networks, Image Classification.

Abstract: Nowadays, Convolutional Neural Network (CNN) architectures are widely used to distinguish animal species.

For example, they are used to differentiate between various types of sheep, dogs, fish, and so on. This greatly

assists people in identifying their species and assessing their value. After all, it is challenging for individuals

to differentiate these animals' species without extensive relevant experience and expertise. Although

Deoxyribonucleic Acid (DNA) testing can be used for identification, it is time-consuming and costly, making

it impractical. Utilizing machine learning methods for differentiation saves a significant amount of time and

effort. However, different CNN architectures have distinct focuses and functionalities. This study compares

the differences between Visual Geometry Group (VGG)16 and EfficientNetB0 by classifying cat breeds. The

primary method is to train models using these two CNNs and then compare their performance, focusing on

their accuracy, computational efficiency, and generalization capabilities. This study reveals the strengths and

weaknesses of these two models, enabling you to understand which neural network is more suitable for use.

1 INTRODUCTION

Today, international organizations have confirmed

that there are more than 70 cat species in the world.

Even a cat lover cannot recognize all cat species

(Zhang, 2020). The task of identifying cat breeds is

not only a trivial pursuit for cat lovers, it also holds

significant importance in various fields such as

veterinary science, animal conservation, and even in

the pet industry (Ramadhan, 2023).

With the advent of deep learning, the accuracy

and efficiency of breed identification have seen

remarkable improvements (Shrestha, 2019). Deep

learning, particularly convolutional neural networks

(CNNs), has revolutionized the way complex image

classification tasks are approached, offering a

powerful tool for distinguishing between different

species and breeds (Li, 2021).

This paper presents a comparative analysis of two

prominent CNN architectures, including VGG16 and

EfficientNetB0, in the context of cat breed

classification. This work trains and evaluates these

models on a dataset of cat images, focusing on their

performance metrics such as accuracy, computational

efficiency, and generalization ability. The goal is to

not only determine which model performs better in

classifying cat breeds but also to understand the

underlying reasons for their performance differences.

2 METHOD

2.1 Dataset

In this study, the dataset comes from Kaggle, named

"Cat Breeds Dataset" (Ma, 2019). This dataset

consists of hundreds of thousands of high-quality

images, covering sixty cat breeds. Because there are

too many cat breeds in this dataset. To facilitate

processing and uploading and save training time, this

work selected the five most common cat breeds for

research, including Calico, Persian, Siamese,

Tortoiseshell, and Tuxedo. The final filtered dataset

has about 14,000 images.

2.2 Data Preprocessing

To ensure a proper balance between training and

validation, this work randomly split the dataset into

training and validation sets in an 8:2 ratio, where 80%

of the data is used for training and 20% is used for

validation. The validation set helps evaluate the

238

Xu, H.

Comparative Analysis of VGG16 and EfﬁcientNet for Image-Based Cat Breed Classiﬁcation.

DOI: 10.5220/0013297100004558

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Modern Logistics and Supply Chain Management (MLSCM 2024), pages 238-242

ISBN: 978-989-758-738-2

performance of the model on unseen data and more

accurately reflects the generalization ability of the

model rather than just its performance on the training

data. By using the validation set, this work can detect

whether the model is overfitting the training data. If

the model performs well on the training set but poorly

on the validation set, it may indicate overfitting. By

fixing the random seed, the split is guaranteed to be

reproducible, which is very useful for debugging and

comparing model performance.

2.3 Data Augmentation

To enhance the model's learning capabilities, this

work implemented a series of data augmentation

techniques. Initially, pixel values were normalized

from [0, 255] to [0, 1] to streamline model processing

and expedite training convergence. The data

augmentation included random rotations up to 40

degrees and translations of up to 20% in width or

height, simulating varied object positions and

orientations. Shearing was applied to introduce

complex deformations, while random scaling up to

about 20% helped the model recognize objects of

different sizes as the same category. Horizontal

flipping of images was also performed to expand the

training dataset. Additionally, any blank spaces

created by these transformations were filled with the

nearest pixel value to preserve image integrity.

Consistent with the training data, the validation set

was normalized to maintain uniform data formatting,

which is essential for effective model training and

evaluation (Xu, 2023).

2.4 Model Architecture

VGG16 is a very early CNN model. When deep

learning became popular, it was the best neural

network architecture at the time (Simonyan, 2014).

So, it is a very good control group, reference group.

It deepens the network by stacking 3*3 convolutional

layers. Its reference volume is relatively large, about

138Millions, and its design is relatively simple. So,

for some basic image classification tasks, its

performance is still relatively good. But it may be a

bit difficult to distinguish the types of cats, because it

needs to capture more complex and subtle features,

and vgg16 is not as good as the new architecture in

this regard.

EfficientNetB0 is a new type of neural network

architecture (Tan, 2019). Its design has been

optimized and is different from VGG16. It balances

depth, width and resolution through compound

scaling technology, greatly improving accuracy. It

has a low reference size, only about 5.1 million

parameters. So, from a design point of view, it is

lighter and more efficient than VGG16. It is stronger

than VGG16 in solving complex image classification

tasks. This study is the results and performance

analysis of the training model to compare the

performance of VGG16and EfficientNetB0 when

used to classify cat breeds. Their performance is

analyzed and compared by using prediction rate,

readiness rate, F1 score and confusion matrix (Hossin,

2015).

3 EXPERIMENT AND RESULTS

3.1 Training Details

This study used the following training techniques to

optimize the model. They are ModelCheckpoint,

EarlyStopping, and ReduceLROnPlateau.

ModelCheckpoint is a callback function that helps

users to save the best performing model on the

validation set during training. EarlyStopping is a

callback function used to stop training early (Yao,

2007). If the performance on the validation set (e.g.,

validation loss) does not improve within a certain

number of epochs, training will end early to avoid

overfitting. ReduceLROnPlateau is a callback

function used to reduce the learning rate when the

model performance does not improve. It

automatically reduces the learning rate when the

validation set performance does not improve within a

certain number of epochs. This helps the model adjust

weights more smoothly when it is close to the optimal

solution.

The model needs to be trained in two phases. The

last few layers were selectively unfrozen for fine

tuning after freezing the convolutional layer at the

first stage, by freezing convolutional layers one can

be sure that it has learned general low-level features

and simple objects. This prevents overfitting. During

the second phase, the author unfrozes most of these

last layers only for some fine-tuning tasks specify

ones to learn new features from this data while

preserving all other learned general features

previously. Applying a low learning rate, fine-tuning

weighs the end layers to smoothly adjust them

through while training on those new tasks.

3.2 Result Comparison

To compare the performance of the two trained

models, this work evaluated the models in various

ways. In addition to accuracy, it also includes

Comparative Analysis of VGG16 and EfﬁcientNet for Image-Based Cat Breed Classiﬁcation

239

precision, recall, F1 score and ROC curve. These

indicators can help people better evaluate the

performance of the model and compare the models.

This work set up a test set of 2495 images, which are

classified into five cat breeds: 'Calico', 'Persian',

'Siamese', 'Tortoiseshell', 'Tuxedo'. Then use this

dataset to test the performance of the two models. The

performance is of VGG16 and EfficientNet is

demonstrated in Table 1 and Table 2, respectively,

with their confusion matrixes in Figure 1 and Figure

Table 1: Performance comparison using VGG16 model.

Precision

Recall

F1-score

Calico

0.753

0.842

0.795

Persian

0.871

0.934

0.901

Siamese

0.910

0.874

0.892

Tortoiseshell

0.889

0.768

0.823

Tuxedo

0.909

0.896

0.902

Accurac

0.863 0.863 0.863

Macro av

0.866 0.863 0.863

Wei

hted av

0.866 0.863 0.863

Table 2: Performance comparison using EfficientNetB0.

Precision

Recall

F1-score

Calico

0.858

0.898

0.878

Persian

0.911

0.988

0.948

Siamese

0.956

0.962

0.959

Tortoiseshell

0.945

0.790

0.860

Tuxedo

0.934

0.960

0.947

Accurac

0.919 0.919 0.919

Macro av

0.921 0.919 0.918

Wei

hted av

0.921 0.919 0.918

Figure 1: Confusion matrix result of VGG16 model (Figure

Credits: Original).

Figure 2: Confusion matrix result of EfficientNetB0 model

(Figure Credits: Original).

For accuracy, the VGG16 model achieves 86.25%

and EfficientNetB0 91.94%. It is obvious that

EfficientNetB0 model surpasses VGG16 for general

accuracy improvement around 5.69% VGG16

weighted average precision: 86.62%EfficientNetB0

weighted average precision: 92.08% Overall

EfficientNetB0 has a better accuracy in almost all

categories and fewer false positives than VGG16.

Moreover, VGG16 has the weighted average

precision of 86.25% and EfficientNetB0, it is

increased to the weighted average recall of 91.94%.

EfficientNetB0 shows a higher recall over VGG16 in

all categories. VGG16 has a weighted average of

86.26% F1 score, and EfficientNetB0 got that number

up to 91.83%. EfficientNetB0 has a better balanced

between precision and recall compared to VGG16

with the higher F1 score. If analyzed by category,

EfficientNetB0 has higher accuracy, precision, recall,

and F1 score than VGG16 for every cat breed. From

the confusion matrix, VGG16 had the most difficulty

distinguishing between certain categories, such as

"Tortoiseshell" and "Calico", as seen in the confusion

between the two in the matrix. For example, 68

Tortoiseshell cats were incorrectly classified as

Calico cats. EfficientNetB0 significantly reduced this

confusion, although there were still some

misclassifications (e.g., 59 Tortoiseshells were

classified as Calico). EfficientNetB0 also showed

better overall performance in identifying the

"Siamese" and "Tuxedo" breeds, with fewer

misclassifications than VGG16.

MLSCM 2024 - International Conference on Modern Logistics and Supply Chain Management

240

Figure 3: ROC curve of VGG16 model (Figure Credits:

Original).

Figure 4: ROC curve of EfficientNetB0 model (Figure

Credits: Original).

As demonstrated in Figure 3, the AUC values of

the VGG16 model show that, except for Calico, all

other categories have scores very close to 1.0 or they

have high accuracy in this type, which tells that in

most cases, VGG16 predicts animals as predicted,

especially the Persian and Siamese cat types, which

are even close to 100% accurate. In addition, as

displayed in Figure 4, EfficientNetB0 has better AUC

than VGG16 overall, with an AUC of 1.00 for Persian

and Siamese cats. EfficientNetB0's Tuxedo is equal

to 1 with VGG16's Calico and Tortoiseshell. Of

course, intuitively, it could be observed that the curve

of EfficientNetB0 is closer to the y-axis, and the

curves of EfficientNetB0 are more concentrated with

each other.

4 DISCUSSIONS

From the above experimental results, pictures and

analysis, it could be observed that although VGG16

is a classic deep learning model and performs well in

multiple classification tasks, its performance is

obviously inferior to EfficientNetB0 in this

experiment. This phenomenon may be related to the

fact that VGG16 has too many parameters and has not

been optimized. Its huge fully connected layer design

is prone to overfitting when processing relatively

small data sets, while increasing training time and

computing resource consumption. In contrast,

EfficientNetB0 is more sophisticated in design and

uses compound scaling. While maintaining a low

parameter volume, it can still achieve higher

classification accuracy. Its performance on the three

varieties of Persian, Siamese and Tuxedo is

particularly outstanding, with an AUC of 1.00, almost

eliminating misclassification. This shows that

EfficientNetB0 not only has stronger discrimination

ability when processing complex visual features, but

also can better adapt to task requirements when the

distribution differences between categories are large.

In terms of overall performance, EfficientNetB0

outperforms VGG16. But in fact, in these five

categories, the AUC of EfficientNetB0 is only

slightly higher than that of VGG16. This may be

related to the selected dataset and the category of cats.

The quality of the dataset in this study is not high, the

images are not complex enough, and there are fewer

fine features. And the five cats selected in this study

look very similar in visual features. Therefore, the

model sometimes misclassifies them.

Future work. This study can still be further

improved. It could be found that some higher quality

datasets to train the model. Works can also use some

better data enhancement methods when processing

data. These can further improve the performance of

the model and its ability to accurately recognize.

Secondly, more categories of cats could be introduced

for research. Then more results and data will make the

research more convincing. Third, several different

neural network architectures could be applied in the

research, obtain the results for analysis and

comparison, to improve the level and quality of the

research.

5 CONCLUSIONS

This study uses two neural network models, VGG16

and EfficientNetB0, to classify cat breeds.

Comparative Analysis of VGG16 and EfﬁcientNet for Image-Based Cat Breed Classiﬁcation

241

EfficientNetB0 shows higher values in key indicators

such as precision, recall, F1-score, and AUC values,

especially in the classification of Persian, Siamese,

and Tuxedo breeds, because its perfect classifier

reaches 1.00. In addition, the micro-average AUC of

EfficientNetB0 is 0.99, which is significantly higher

than 0.98 of VGG16.

Compared with VGG16, EfficientNetB0 is more

efficient and has significantly fewer parameters,

which not only leads to better performance, but also

lower computational complexity and training time

compared with previous architectures. Based on the

comparative analysis of ROC curves, EfficientNetB0

shows stronger classification capabilities, especially

in the case of reducing misclassification. The result is

that compared with VGG16, EfficientNetB0 is a more

advantageous model in the task of cat breed

classification, and its excellent classification

performance and efficient computational

performance make it have broad application

prospects in similar image classification tasks.

Overall, this study reveals the significant

advantages of EfficientNetB0 in performance and

computational efficiency through data comparison

between EfficientNetB0 and VGG16, and draws

some reasonable analysis, results and conclusions.

This also provides a good reference for those who

attempt image classification tasks in the future.

REFERENCES

Hossin, M., & Sulaiman, M. N. 2015. A review on

evaluation metrics for data classification

evaluations. International journal of data mining &

knowledge management process, 5(2), 1.

Li, Z., Liu, F., Yang, W., Peng, S., & Zhou, J. 2021. A

survey of convolutional neural networks: analysis,

applications, and prospects. IEEE transactions on

neural networks and learning systems, 33(12), 6999-

7019.

Ma. 2019. Cat breeds dataset. Kaggle. URL:

https://www.kaggle.com/datasets/ma7555/cat-breeds-

dataset. Last Accessed: 2024/09/14

Ramadhan, A. T., & Setiawan, A. 2023. Catbreedsnet: An

Android Application for Cat Breed Classification Using

Convolutional Neural Networks. Jurnal Online

Informatika, 8(1), 52-60.

Shrestha, A., & Mahmood, A. 2019. Review of deep

learning algorithms and architectures. IEEE access, 7,

53040-53065.

Simonyan, K., & Zisserman, A. 2014. Very deep

convolutional networks for large-scale image

recognition. arXiv preprint arXiv:1409.1556.

Tan, M. 2019. Efficientnet: Rethinking model scaling for

convolutional neural networks. arXiv preprint

arXiv:1905.11946.

Xu, M., Yoon, S., Fuentes, A., & Park, D. S. 2023. A

comprehensive survey of image augmentation

techniques for deep learning. Pattern Recognition, 137,

109347.

Yao, Y., Rosasco, L., & Caponnetto, A. 2007. On early

stopping in gradient descent learning. Constructive

Approximation, 26(2), 289-315.

Zhang, Y., Gao, J., & Zhou, H. 2020. Breeds classification

with deep convolutional neural network.

In Proceedings of the 2020 12th international

conference on machine learning and computing. 145-

151.

MLSCM 2024 - International Conference on Modern Logistics and Supply Chain Management

242