Received 15 June 2023, accepted 10 July 2023, date of publication 24 July 2023, date of current version 28 July 2023.

Digital Object Identifier 10.1109/ACCESS.2023.3297440

Hybrid Deep Learning Algorithms for Dog Breed

Identification—A Comparative Analysis
B. VALARMATHI 1 , N. SRINIVASA GUPTA 2 , G. PRAKASH 3, R. HEMADRI REDDY 4,

S. SARAVANAN 5 , AND P. SHANMUGASUNDARAM 6

1 Department of Software and Systems Engineering, School of Information Technology and Engineering, Vellore Institute of Technology, Vellore, Tamil Nadu
632014, India
2 Department of Manufacturing Engineering, School of Mechanical Engineering, Vellore Institute of Technology, Vellore, Tamil Nadu 632014, India
3 Department of Database Systems, School of Computer Science and Engineering, Vellore Institute of Technology, Vellore, Tamil Nadu 632014, India
4 Department of Mathematics, School of Advanced Sciences, Vellore Institute of Technology, Vellore, Tamil Nadu 632014, India
5 Department of Electronics and Communication Engineering, Srinivasa Ramanujan Centre, SASTRA Deemed University, Kumbakonam, Tamil Nadu 612001,

India
6 Department of Mathematics, College of Natural and Computational Sciences, Mizan-Tepi University, Mizan Teferi 5140, Ethiopia

Corresponding author: P. Shanmugasundaram ([email protected])

ABSTRACT Deep learning and computer vision algorithms will be applied to find the breed of the dog
from an image. The goal is to have the user submit an image of a dog, and the model will choose one of
the 120 breeds stated in the dataset to determine the dog’s breed. The proposed work uses various deep
learning algorithms like Xception, VGG19, NASNetMobile, EfficientNetV2M, ResNet152V2, Hybrid of
Inception &Xception, and Hybrid of EfficientNetV2M, NASNetMobile, Inception &Xception to predict
dog breeds. ResNet101, ResNet50, InceptionResNetV2, and Inception-v3 on the Stanford Dogs Standard
Datasetswere used in the existing system. The proposed models are considered a hybrid of Inception-v3
&Xception and a hybrid of EfficientNetV2M, NASNetMobile, Inception & Xception. This hybrid model
outperforms single models like Xception, VGG19, InceptionV3, ResNet50, and ResNet101.The authors used
a transfer learning algorithm with data augmentation to increase their accuracy and achieved a validation
accuracy score of 71.63% for ResNet101, 63.78% for ResNet50, 40.72% for InceptionResNetV2, and
34.84% for InceptionV3. This paper compares the proposed algorithms with existing ones like ResNet101,
ResNet50, InceptionResNetV2, and InceptionV3. In the existing system, ResNet101 gave the highest
accuracy of 71.63%. The proposed algorithms give a validation accuracy score of 91.9% for Xception,
55% for VGG19, 83.47% for NASNetMobile, 89.05% for EfficientNetV2M, 87.38% for ResNet152V2,
92.4% for Hybrid of Inception-v3 &Xception, and 89.00% for Hybrid of EfficientNetV2M, NASNetMobile,
Inception &Xception. Among these algorithms, the Hybrid of Inception-v3 &Xception gives the highest
accuracy of 92.4%.

INDEX TERMS EfficientNetV2M, hybrid of EfficientNetV2M, hybrid of inception and Xception, NasNet-
Mobile, ResNet152V2, VGG19, Xception.

I. INTRODUCTION tool in artificial intelligence applications. Optical Charac-

Nowadays, there is an increasing demand and usage of image ter Recognition and facial recognition are two examples of
classification and verification techniques; the most signifi- computer vision applications. These areas offer impressive
cant technique used for image data classification is a deep results, fueling increasing interest in deep learning. Classify-
learning model known as Convolutional Neural Networks ing images is a field in which deep learning excels. CNN is the
(CNN). Deep learning is increasingly becoming a crucial most popular deep-learning method for classifying images.
The proposed work is to investigate a variety of Convolu-
The associate editor coordinating the review of this manuscript and tional Neural Network models for classification. The created
approving it for publication was Nuno M. Garcia . algorithm might be used in a mobile or online application.

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
77228 VOLUME 11, 2023
B. Valarmathi et al.: Hybrid Deep Learning Algorithms for Dog Breed Identification

Transfer training, which helps us produce effective outcomes, set for this study. The data set was split into two categories:
was used to create the CNN that categorizes dog breeds. training and testing, with the training set being used to train
Convolutional Neural Networks (CNNs) are made up of the established model. Using more widespread pet dogs as
neurons with adjustable biases and weights. The dot product samples, it was suggested by the authors a yolov3-based
is computed by each neuron using input data. Convolution pet dog categorization model. Instead of using the pet dog’s
Neural Network architectures differ from other neural net- general characteristics, the model chose the face of the pet
works in that itrelies on actual images as inputs, unlike other dog and marked the dog category by detecting facial features,
neural network types. This enables the inclusion of partic- similar to face detection. The trials’ findings demonstrated
ular components in the design. CNN lower the number of that this approach was capable of rapidly and correctly locat-
observable variables that characterize a network. Single-layer ing the position of a pet dog’s face, and it resolved the issue
neurons act independently from prior layers. of pet dog detection and categorization.
A convolutional neural network architecture from the In this paper [2], the authors studied how animal identifi-
Inception family called Inception-v3 employs an additional cation in veterinary practice was managed through machine
classifier, factorized 7 × 7 convolutions, and label smoothing learning. Electronic animal health records included digi-
to carry label information further down the network. tal image graphs using image processing and recognition
Deep-separable convolution layers make up the complete technology to identify animals. The authors studied how
architecture of the CNN called Xception. combining ‘‘soft’’ biometrics, such as breed and facial bio-
A type of Convolution Neural Network called NASNet was metrics, could improve dog identification. The researchers
found through research into neuronal design. Normal cells applied transfer learning from GoogLeNet to propose Breed-
and decreasing cells serve as the foundation. Net for breed classification and subsequently to propose
EfficientNet is a CNN architecture; all depth, width, DogNet for identifying individual dogs within the classified
and resolution factors are scaled using a compound coeffi- breeds.
cient method. The EfficientNet scaling technique, in contrast In this article [3], by determining a dog’s breed in a given
to conventional practice, uniformly adjusts network width, image, this work aimed to solve the problem of fine-grain and
depth, and resolution using a collection of fixed scaling coef- multi-class image recognition. One of the sophisticated deep
ficients. learning approaches used in the research system was convo-
Residual Networks, or ResNets, rather than learning non- lutional neural networks. Two distinct networks were con-
referenced functions, learn residual functions about layer structed and assessed using the Stanford Dogs dataset. Con-
inputs.The residual nets enable those layers to match a resid- volutive neural networks’ application and evaluation were
ual map instead, assuming that each stacked layer exactly demonstrated using a software system. It had a central server
matches a desired underlying map. ResNets stack residual and a mobile client with resources and tools for online and
blocks together to build networks, such as ResNet-50, which offline neural network analysis. Two distinct convolutional
has 50 layers. neural network architectures were presented: the Inception-
A deep convolutional neural network is generally chal- ResNet-v2 deep architecture and the NASNet-A mobile
lenging to train from the start. Because size and depth data architecture. Deep Inception-ResNet-v2 model outperformed
set relevant to the neural network are rare, using transfer even the smaller, mobile-friendly CNN, with results that were
learning as a feature extractor is optimal. A model wasalready still encouraging.
trained on a large data collection. Transfer learning is a crucial This study [4] provided two models for categorizing dogs
component of deep convolutional neural networks that offers into different breeds. Due to the increasing difficulty of clas-
solutions to these issues. In computer vision, transfer learning sifying dogs and the fact that these classifications were based
is defined as using a pre-trained model. A pre-trained model on deep learning, forming the two models that provide differ-
must have been trained on a sizable benchmark dataset to ent levels of accuracy at both ends requires a fully defined
handle an issue similar to ours. The magnitude of the new data set. Since every model was periodically subjected to
dataset and how closely it resembles the existing dataset predictions, the researchers encountered numerous function-
play a significant role in determining transfer learning. So, ing levels during the investigation that weren’t considered in
before using transfer learning, a few circumstances shall be earlier research. The essential concept of transfer learning,
considered. The efficacy of the CNN is reduced by overfitting which dealt with the data augmentation technique and its
when the new dataset is smaller but contains the same data as capacity to increase the size of the data set, is also built upon
the old dataset. If the new dataset is large and has content by their approach. Afterward, accuracy levels were matched
comparable to the old data, the model can be refined using or compared with both models to establish a comparison for
the entire network. both models. A detailed procedure was also used to classify
the data. The comparison between Inception V3 and VGG16
II. LITERATURE SURVEY was offered in the publication. Observations showed that
In this research [1], to categorize and identify pet dogs’ faces, Inception V3 offered an accuracy of 85, whereas VGG16
the researchers offered an improved Yolov3 model. Eight provided a much lower accuracy of 69 than the Inception V3
distinct breeds of pet dogs were used to construct the data model.

VOLUME 11, 2023 77229

 B. Valarmathi et al.: Hybrid Deep Learning Algorithms for Dog Breed Identification

In this article [5], deep learning made it possible to train proper breed(s) were identified using image processing algo-
algorithms (models) that could categorize and forecast data rithms for breed classification because the dogs might be
based on the knowledge that was extracted (learned) from cross-breed ancestors. It would be necessary for straightfor-
the raw data. Convolutional Neural Networks were one ward breed-based dog categorization, and it could demon-
method that was commonly employed for image categoriza- strate how unreliable or trivial breed identification by the
tion and detection. In this research, the authors discussed naked eye is. The authors could study and perform animal
dog breed/type identification after providing a CNN-based recognition using image processing, including sheep, cat-
method for dog detection in potentially complex images. The tle, and others. VGG Net was an extensive collection of
findings had a 64% accuracy rate for 120 additional, less pre-defined CNN demonstrating its ability to utilize/prepare
common dog varieties and a breed classification accuracy pictures on most subjects accurately.
of about 85% for a group of 50 dog breeds. A big data This article [9] used Convolutional Neural Networks
processing infrastructure using a variety of GPUs and an iOS specifically for dog breed detection. Despite being an effec-
application supporting image classification techniques were tive method, convolutional neural network classification still
used. To enhance the integrity of the data, several preparation has a few flaws. For training, Convolutional Neural Networks
methods were employed. require many images and much time to get better classifi-
In this study [6], there were many different species and cation accuracy. The authors employed transfer learning to
organisms in the world today. This highlights how important get around this lengthy period. Learning in computer vision
it was to classify different tangible objects. Determining the refers to training the CNN using a pre-trained model. A pre-
similarities between distinct classes also became extremely trained model was trained to solve a classification problem
important in light of the continuing genetics and evolution comparable to the onethe researchers had using transfer learn-
study being conducted by scientists worldwide. The exper- ing. In this research, the authors trained over 1400 photos
iment that provided the basis for this essay involved classi- encompassing 120 dog breeds using various pre-trained mod-
fying different canine breeds using a CNN (Convolutional els, including VGG16, Xception, and InceptionV3. Then the
Neural Network). This algorithm will identify an estimated bottleneck features were used in these pre-trained models to
breed if a canine image is found. The type of dog that most identify bottlenecks. Last, Logistic Regression, a multiclass
closely resembled a human, if one was given, was identified. classifier, was used to determine the dog breed from the
The authors created a pipeline to handle photos from the photos. It achieved 91%, 94%, and 95% validation accuracy
actual world. The dog breed classifier did an excellent job, for the various pre-trained models VGG16, Xception, and
with extremely good accuracy. InceptionV3.
In this article [7], the researchers used cutting-edge models The dog was one of the most common domesticated ani-
on Imagenet data sets. To extract the feature from the data mals. Having so many dogs led to several issues, including
set used to identify dog breeds, the pre-trained model and population control, decreased rabies outbreaks, vaccination
learned weights were used. After that, data augmentation control, and formal ownership. Currently, there are about
and fine-tuning were used to improve the performance of 180 distinct canine breeds. Each canine breed had distinctive
their test’s breed categorization accuracy. The performance characteristics and health problems. To deliver the appropri-
of the proposed methods was compared with the GoogleNet, ate therapies and training, it was essential to identify individ-
DenseNet-169, DenseNet-121, and ResNet-50 models from uals and their breeds. The article illustrates the classification
the most current Image-Net dataset. Their respective test approaches for dog breed classification using two image
accuracies were 82.08%, 84.01%, 85.37%, and 89.66%, processing techniques. 1) Two instances of conventional tech-
demonstrating the proposed method’s superior performance niques were the Histogram of Oriented Gradient and the
to earlier efforts on Stanford dog breed datasets. Stanford dog Local Binary Pattern (LBP). 2) The deep learning-based
breed datasets presented a modified approach to cutting-edge strategy utilizing transfer learning and Convolutional Neural
networks like ResNet, DenseNet, and GoogleNet. Due to the Networks (CNN). The outcome demonstrated that their CNN
limited training datasets, data augmentation and fine-tuning model that had been trained better performed when classify-
were conductedto improve the test set experiments’ accuracy. ing dog breeds. Compared to 79.25% accuracy when using
In this paper [8], the character of an animal like a dog the HOG descriptor, it achieved 96.75% accuracy [10].
has changed significantly from earlier generations due to With improved techniques, image classification made sig-
extensive breeding or cross-breeding. Contrary to eye-only nificant progress and improved accuracy. However, there was
recognition, image processing for breed analysis allowed the a ton of room for improvement regarding fine-grained classi-
most accurate prediction of the exact outcome or results. fication. The different animals could all be recognized from
Breed analysis and identification were made using the ADA the image, but it was more challenging to determine the breed
boosting methodology. By combining numerous weak clas- of each animal. This essay aims to advance the classification
sifiers, ADA Boosting produced a robust classifier. The of animal breeds. Several pre-trained deep learning models
authors employed image processing classification to distin- were trained and tested using the standard Stanford dog breed
guish between the many dog breeds. It accurately predicted dataset. The pre-train network was tweaked, and the results
the dominant breed or breeds present in the canine. The were compared. The training process data recorded while

77230 VOLUME 11, 2023

B. Valarmathi et al.: Hybrid Deep Learning Algorithms for Dog Breed Identification

fine-tuning the AlexNet model was displayed in graph and The authors used InceptionV3, MobileNet, VGG- 16,
tabular forms. Comparative analysis was performed using & Xception algorithms, and Stanford Dogs (ST) Dataset,
each network that has recorded data, and the results are dis- Columbia Dogs (CU) Dataset & Flickr-dog Dataset were used
played. DenseNet201 reported a testing accuracy of 87.15%, in this paper [14]. CU and ST were used to classify canine
GoogleNet reported a testing accuracy of 81.53%, AlexNet breeds, and Flickr-dog was considered for dog identification.
reported a testing accuracy of 84.35%, and ResNet50 reported The identification rate of dogs was 78.09% without using
a testing accuracy of 90.12%. The final three layers of each ‘‘soft’’ biometrics, but by using a decision network to com-
network were modified to produce these results. The study bine ‘‘soft’’ biometrics, the identification rate could reach
could be expanded by considering more models and experi- an accuracy of 84.94%. The suggested strategy, which relied
menting with layer modifications [11].Classifying images of solely on CNNs, produced average accuracy gains of 6.2%.
dogs by breed was difficult, making it a problematic picture The identification process using a ‘‘fusion’’ on a classifier
classification task. In this study [12], a Convolutional Neural decision achieved approximately 11.2% higher accuracy.
Network (CNN) and transfer learning model-based Android This paper [15] used Xception and Multilayered Percep-
application was created that analyses images to identify a tron (MLP) algorithms. The dog breed dataset was derived
dog’s breed. The Android application allows users to capture from a Kaggle contest dog breed identification. The dataset
or upload dog photos. The features required for testing were comprised 120 unique dog breeds and 10,222 images of
then extracted after the picture underwent pre-processing. dogs. LogLoss and Balanced Accuracy; the optimal model
Based on transfer learning and CNN, dog breed predictions produced an accuracy of 0.5480, or 54.80%. The results were
were made. The model was taught using Stanford’s standard obtained using only three splits. The accuracy achieved by
dog dataset, and it had a 94% accuracy rate when tested this model was not satisfying. The number of splits could
against actual data. The intended work had been effectively be raised, allowing the model to train more and improving
designed, implemented, and tested. The authors had created prediction accuracy. Other approaches may be tried that are
a simple Android application that allowed users to submit or more accurate at predicting dog breeds. Additionally, various
choose an image to determine a dog’s breed. There was a very breed combinations could be trained to detect variations in
short waiting time because it worked without an internet link accuracy.
and gave the answer right away. In this study it was shown In this article [16], the authors put forth a brand-new frame-
how to build a dog breed recognition model and deploy it work model that was referred to as SC-MPEM (Supervised
on an Android device using pre-trained models. The appli- Clustering Using Multi-Part CNN and EM), which makes use
cation’s size could be reduced in the future with the help of of the Inception v3 network for training and the YOLOv3
enhancements. Additionally, the model that correctly predicts for discriminative part detection. There were four distinct
photographs captured in another way could be improved. benchmark datasets used, like the Oxford-IIIT Pet dataset
Deep learning neural networks have recently gained pop- (OD), the Columbia Dogs with Parts (CD), Stanford Dogs
ularity and are used in various industries, including finance (SD), and camera trap pictures from the Snapshot Serengeti
and healthcare, travel, media, retail, etc. The methodology for datasets. It has been demonstrated that deep CNNs trained
optimizing CNN, as used in the Stanford dataset of canine under supervision on a sizable and diverse dataset extract
breeds, was presented in the current work [13]. The deep superior features than most traditional methods, even for
neural network, which had weights and biases, was com- unsupervised tasks. The novel yet straightforward proposed
parable to the Convolution Neural Network. The specific approach outperformed other state-of-the-art models. To fur-
features or patterns contained in the original data were pre- ther increase the stiffness of the training dataset, images
dicted using CNN filters. Modern technologies frequently of the animals could be embraced in various poses (facing
employed trained Convolutional Neural Networks that had away from the camera) under various lighting circumstances
been fine-tuned. Many refined transfer learning techniques (day and night). The animal detection algorithms could use
are in use now. Inception-ResNet-V2 was implemented over thermal images to prevent illumination issues with visible
the dataset in this application. Only mentioning the dog’s images.
breed was insufficient; it was also crucial to mention its This paper [17] used Convolutional Neural Networks
origin, color, height, weight, longevity, health, training, and with modals like InceptionResNet V2 and InceptionV3.
other traits specific to each breed. Web scraping or web data Also,the researchers used the Stanford Dogs dataset, contain-
extraction was used to retrieve data from websites. As a result, ing 120 unique dog breeds and 10222 dog images for training,
web scraping could complete the same activity faster than and 10357 images for testing.
manually copying data from websites because the process This study [18] used a convolutional neural network
was automated. Reference websites like Wikipedia and Dog and TensorFlow model called MobileNet for mobile and
Breed List were used in this application to obtain crucial embedded mobile application visions. A self-made dataset
data that was then rendered utilizing web scraping with a of 1000 dog images were used. The algorithm used here
respectable user interface and user experience. This paper (CNN) gave good accuracy for all the tested datasets. Transfer
gave numerous traits and crucial information about the dog learning made an excessive decision by combining a prebuilt
based on the supplied image result. model with the model developed in this research. Analysis

VOLUME 11, 2023 77231

 B. Valarmathi et al.: Hybrid Deep Learning Algorithms for Dog Breed Identification

was done only for dogs, but planning to extend this project so In this study [3], the authors offered a fine-grained, multi-
that other animals can be identified. class image identification challenge to identify a dog’s breed
In this article [19], Convolutional Neural Networks (CNN) in a given image precisely. The presented system used mod-
with pre-trained models like ResNet50 were used in this ern deep learning methods, including Convolutional Neural
paper [19]. The authors used dog images of 133 different Networks. Two different networks were trained and assessed
dog breeds. 8,351 dog images and 13,233 human images using the Stanford Dogs dataset. A software system demon-
in total were used to identify dogs only when they were strated the application and assessment of Convolutional Neu-
confirmed to be canines, allowing the system to determine ral Networks. It included a central server, a mobile client,
whether the given image more resembled a person or the and parts and modules for online and offline neural network
closest predicted dog breed. 82.7% accuracy using CNN with analysis. The Inception-ResNet-v2 deep architecture and the
ResNet50 algorithm (transfer learning). Of all the algorithms, NASNet-A mobile architecture were two different convolu-
it was discovered that this one had the highest accuracy. tional neural network designs that had been introduced. The
12% of dogs were misclassified as people, which could be designs were evaluated using a particular image classification
decreased by using a bigger dataset. challenge: identifying dog breeds. The pre-trained networks
In this paper [20], CNN with three representative mod- were adjusted using the Stanford Dogs dataset. The findings
els, VGG16, Inception V3, and Xception was used in this were encouraging even for the smaller, mobile-friendly CNN,
work [20]. Stanford Dog data set with 20,580 dog images with only 10% less accurate than the deep Inception-ResNet-
for 120 different breeds of dogs was used in this work. The v2 mode.
accuracy score of Xception was 99%, Inception was 94%, and In this paper [2], the authors examined how combining
VGG16 was 85%. One drawback was that learning was too ‘‘soft’’ biometrics, such as canine breed and face biometrics,
slow; when the dataset grew bigger, VGG16 + LR (Logistic could enhance dog identification. The proposed BreedNet
Regression) did not perform as well. The learning rate could was used to classify breeds, and the proposed DogNet was
be improved by considering other models. used to recognize specific dogs within the classed breeds,
In this study [21], a deep learning-based technique for using transfer learning from Google LeNet. To categorize dog
identifying dog breeds using face photos was demonstrated breeds and then recognize specific dogs using photographs,
in the study. To increase accuracy, the suggested approach the ‘‘coarse-to-fine technique’’ and transfer learning were
combines pre-trained CNNs with the transfer learning tech- used. The proposed BreedNet’s breed categorization accu-
nique. Three CNN models—MobilenetV2, InceptionV3, and racy was comparable to the highest outcomes previously
NASNet—were examined in the experiments. Each model reported. Breed categorization first reduces the search space
was developed using training sets of images enhanced with for additional canine identification by identifying the top-k
random noise, rotation, and other effects. With a rotation potential breeds given a probe picture of a dog. The BreedNet
image training set, the NASNet model achieved the highest learned for breed categorization was converted to DogNet
accuracy of 89.92%. The rotation might help with picture for canine identification using transfer learning, allowing the
alignment because the model primarily concentrated on the same CNN architecture at both the ‘‘coarse’’ and ‘‘fine’’
center of the images. With a classification accuracy of more stages. Comparing the suggested method to previous works,
than 80% in all scenarios, the suggested approach could accuracy was 15% higher.
deliver a promising performance. It may be highly accurate In this article [23], the authors used Convolutional Neural
with enhanced datasets like rotation and translation. Networks to categorize dog breeds in a highly exact way.
This paper [22] classified different dog breeds using CNN. It fell under the domain of fine-grained image classifica-
If a canine image were provided, the algorithm would search tion problems, in which inter-class variances were modest,
for the breed of the dog and similarities in the breed’s features. and one little area of the image analyzed often makes the
If a human image was provided, it was determined which difference in categorization. ImageNet classes could have
facial features would show in a dog and vice versa. considerable inter-class variances, making it easier to cat-
In this paper [6], the classification of different canine egorize accurately. It was intended to train and categorize
breeds was done using the convolutional neural network. dog breeds using a Convolutional Neural Network frame-
If an image of a dog is discovered, this algorithm will esti- work. It began by employing CNNs based on the LeNet and
mate the breed. An associated dog breed was determined GoogLeNet architectures.
if a human image was given. The researchers developed a In this article [24], the researchers found that the two
pipeline for processing real-world photos. This method could networks used, VGG-16 and DenseNet201, could identify
be improved by teaching it to distinguish between humans humanly perceptible patterns when fine-tuned on the Stan-
and dogs. Accuracy might be increased even further through ford Dogs dataset. Even though there was over-fitting in
data augmentation. The network could identify features inde- both networks, the necessary measures were taken to avoid
pendent of orientation or scale due to data augmentation. It is and lessen its effects. Their results were presented and ana-
obvious that using transfer learning to create a convolutional lyzed to show that both networks could still identify patterns
neural network was much more accurate than creating one despite the over-fitting. The authors examined both networks’
from the start. response maps (or feature maps) to identify breed-specific

77232 VOLUME 11, 2023

B. Valarmathi et al.: Hybrid Deep Learning Algorithms for Dog Breed Identification

characteristics. Combining their knowledge of networks and TABLE 1. Dog breed name with ID.
35 features, the experts could interpret the networks given
Locke’s theory of ideas and words. Although the article’s
authors rejected the idea that these networks were conscious,
they were a good match for a Lockean interpretation.
In this article [25], Multiple Microsoft Kinect v2s were
used to record a range of dog breeds, and a motion capture
system was also used to acquire the 3D ground truth skeleton.
The topic of 3D canine poses estimation from RGBD pictures
was the focus of their study. Using this information, several
false RGBD images were produced. The authors used prior
models of form and pose to constrain a stacked hourglass
network trained to forecast the positions of 3D joints. Their
model was tested on fictitious and real RGBD images, com-
paring their findings to previous studies that fitted canine
models to images.
This research used a dataset of 70 dog breeds to train and
test transfer deep learning algorithms [26]. The dataset was
statistically stable, containing approximately 100 images of
each category of dog breeds. Then, different deep learning
methods were used, such as Convolutional Neural Network,
InceptionNet, InceptionResNet, VGG16, ResNet, DenseNet,
and etc. The results from algorithm training and testing were
compared based on measures like accuracy, precision, recall,
and area under the curve.
In this study [27], the authors provided instructions for
building a residual neural network to categorize dog breeds
according to a sporting category. The system’s objective FIGURE 1. Sample Data from the dataset.
was to make it easier for people to recognize the different
canine breeds. The Tsinghua Dogs dataset provided the five
Out of 10,222 images, one sample image is shown in
distinct dog breed types used. The same setup was used to
Figure. 1. Twenty images with corresponding labels of
evaluate ResNet 50 and ResNet 101, two CNN implementa-
the dogs are shown in Figure 2. The 120 categories of
tions. Based on the study’s results, ResNet 101 demonstrated
10,222 image class distribution with count value (i.e.) an
improved macro-average f1-score outcomes while maintain-
image count of each dog breed in the dataset are shown in
ing high accuracy. A ResNet 50 f1-score was 84%, while a
Figure 3. Out of 120 categories of dog breed names with
ResNet 101 f1-score was 86%.
numbers, the first five samples are 0: ‘affenpinscher’, 1:
To classify acute lymphoblastic leukemia (ALL) using
‘afghan_hound’, 2: ‘african_hunting_dog’,3: ‘airedale’, 4:
microscopic white blood cell images [28], the authors sug- ‘american_staffordshire_terrier’, and 5: ‘appenzeller.’
gested a hybrid Inception v3 XGBoost model for their work.
The XGBoost model served as the classification and Incep-
IV. PROPOSED WORK
tion v3 as the image feature extractor.
To conduct a comparison of the accuracy values in this paper,
III. DATASET DESCRIPTION & SAMPLE DATA the suggested methods make use of seven distinct algorithms
The dataset used for this research is available at the link like Xception, VGG19, NASNetMobile, EfficientNetV2M,
provided below. ResNet152V2, and two hybrid methods [Hybrid of Incep-
http://vision.stanford.edu/aditya86/ImageNetDogs/ tion &Xceptionand Hybrid of EfficientNetV2M, NASNet-
The Stanford Dogs dataset includes images of 120 distinct Mobile, Inception &Xception] to predict dog breeds. The
canine breeds. This dataset was produced using images and proposed work will evaluate the seven algorithms and deter-
annotation from ImageNet for the purpose of fine-grained mine the most precise and effective. Existing algorithms such
image categorization. Below is a list of what this dataset as ResNet101, ResNet50, InceptionResNetV2 and Inception-
contains. v3 are used.
• 120 categories Figure 4 illustrates the various steps that will be taken
• 10,222 images during the execution of this task to produce the desired result.
• Class labels and bounding boxes are annotations Step 1 (Import Modules): The first step is to import impor-
Out of 120 categories, the sample five dogs’ breed name tant libraries required for the proposed work. The matplotlib
(Boston bull, dingo, Pekinese, bluetick & Golden Retriever) and seaborn libraries are used for graphs. scikit_learn library
and the corresponding ID is shown in Table 1. is used for training and testing splitter. Numpy and pandas’
VOLUME 11, 2023 77233
 B. Valarmathi et al.: Hybrid Deep Learning Algorithms for Dog Breed Identification

FIGURE 2. A few dog images with corresponding labels.

libraries are used to handle image arrays. Tensorflow is used transform images into arrays because it can be mainly used
to utilize the pre-trained models and train our own model. to work with arrays.
Step 2 (Load Dataset): Load the Standford dataset in the Step 7 (Import the Optimizer): The proposed work uses
training folder into the RAM during the runtime. Adam as an optimizer. The adaptive moment estimation
Step 3 (Analyze and Visualize the Dataset): Analyze the provides the foundation for the name Adam. The network
total number of images and the class distribution of the dogs weights are updated during the exercise using this opti-
to identify any gaps in the dataset. Visualizing data by using mization algorithm, another extension of stochastic gradi-
Pandas data frame & functions and plotting graphs. ent descent. The Adam Optimizer continuously updates the
Step 4 (Validate Dataset): Check if the number of labels learning rate for each unique network weight. The Adam opti-
and images is equal. If yes, all images are labeled, and it can mizer is commonly used because of its numerous advantages.
be proceeded further. The algorithm also runs quicker, requires less memory, and
Step 5 (Encode Categorical Classes): Encode categorical requires less tuning than earlier optimization algorithms. It is
classes by assign a unique number to each of the class of the also simpler to implement.
dog breeds. Step 8 (Extract Features): Now, the features of the images
Step 6 (Convertintoarray): To train a deep learning model must be extracted in the form of the NumPy array that
using the characteristics of an image, it is necessary to con- was previously framed. For this, each pre-trained model’s
vert an image into an array. The Python NumPy library can pre-processor will be used to pre-process the images. After

77234 VOLUME 11, 2023

B. Valarmathi et al.: Hybrid Deep Learning Algorithms for Dog Breed Identification

FIGURE 3. The class distribution of dog breed images per class.

VOLUME 11, 2023 77235

 B. Valarmathi et al.: Hybrid Deep Learning Algorithms for Dog Breed Identification

model, feature selection was performed using four models

(EfficientNetV2M, NASNetMobile, Inception-v3, and Xcep-
tion) concurrently and then concatenate the extracted features
into a single NumPy array.
Step 9 (Free Up Resources): Garbage values were collected
to free up some RAM after the feature selection is done.
Step 10 (Model Definition): Now, the proposed model is
defined, including the dropout layer and dense layer with
SoftMax activation function. The model with Adam Opti-
mizer was finally compiled
Step 11 (MODEL TRAINING): After that, fit the model
(train) with the extracted features over 100 epochs and
256 batch size.
Step 12 (Image Prediction): Now, any input image can be
used and extract features from the image and predict using
the trained model.
Step 13 (Accuracy Calculation): The metric used for the
proposed work is accuracy. In this work, it has been calculated
training and validation accuracies for the various models.
Step 14 (Accuracy Comparison): Here, the accuracies
of various models like Xception, VGG19, NASNetMobile,
EfficientNetV2M, ResNet152V2, Hybrid of Inception &
Xception, and Hybrid of EfficientNetV2M, NASNetMobile,
Inception & Xception are compared.
Step 15 (Declare the Best Model): The next stage is to name
the best model based on the highest accuracy achieved after
comparing the accuracy of various models.

V. RESULTS AND DISCUSSION

In the existing system [17], the authors achieved a vali-
dation accuracy score of 71.63% for ResNet101, 63.78%
for ResNet50, 40.72% for InceptionResNetV2, and 34.84%
for Inception-v3. Table 2 also lists the results of the train-
ing and validation accuracy values for different models in
the existing systems. Among these four algorithms, the
Inception-v3 algorithm achieved a low validation accuracy
value of 34.84%, whereasthe ResNet101 algorithmachieveda
high validation accuracy value of 71.63%. The comparative
analyses of the training and validation accuracies for various
models in the existing systems are shown in Figure 5.
The proposed algorithms give a validation accuracy score
of 91.9% for Xception, 55% for VGG19, 83.47% for
NASNetMobile, 89.05% for EfficientNetV2M, 87.38% for
ResNet152V2, 92.4% for Hybrid of Inception-v3 & Xcep-
tion, and 89.00% for Hybrid of EfficientNetV2M, NASNet-
Mobile, Inception & Xception. Among these algorithms, the
Hybrid of Inception-v3 & Xception gives the highest accu-
FIGURE 4. Flowchart of the proposed method. racy of 92.4%.
For the comparative analysis, the proposed model can be
trained with Inception and Xception alone. When trained,
this,feature selection was performed using the models. For an accuracy of 91.4% is given with the Inception-v3 and
the individual model pipeline, it has been used Efficient- an accuracy of 91.9% with the Xception model. The hybrid
NetV2M, NASNetMobile, and ResNet152V2. For the first model of Inception-v3 and Xception predicted 92.4% accu-
hybrid model, feature selection was performed using two racy, greater than that of the Inception and Xception mod-
models (Inception-v3 and Xception); for the second hybrid els found individually. The least accurate model, with an

77236 VOLUME 11, 2023

B. Valarmathi et al.: Hybrid Deep Learning Algorithms for Dog Breed Identification

FIGURE 5. Comparative analysis of the training and validation accuracies for various models in the existing systems.

FIGURE 6. Comparative analysis of the training and validation accuracies for various models in the proposed system.

accuracy of 55%, was the VGG19. Table. 3 displays the The comparative analysis of the accuracy of different algo-
accuracy values from training and validation for the different rithms during training and validation in the proposed system
models in the proposed methodology. is shown in the figure 6. In the proposed hybrid model, the

VOLUME 11, 2023 77237

 B. Valarmathi et al.: Hybrid Deep Learning Algorithms for Dog Breed Identification

TABLE 2. Training and validation accuracies for various models in the work on canine breed prediction. This is an approach that has
existing systems.
promise for upcoming tasks given the success of our hybrid
detection network.
Due to time and CPU limitations, it was found difficult
to conduct many iterations of our technique using Neural
Networks because training each and every layer of the model
is a very time-consuming process. So, in order to train the
complete model and increase accuracy, we anticipate receiv-
ing better GPU and equipment.
We advise further research into neural networks for key
point detection, particularly by training networks with dif-
ferent architectures and batch iterators to see what strategies
might be more effective. In the end, neural network archi-
tectures take a long period to train and iterate, which should
be taken into account for future work. However, compared
to more conventional methods, neural networks are powerful
TABLE 3. Proposed methodology accuracy values from training values classifiers and will improve prediction accuracy.
and validation for the different models.

REFERENCES
[1] C. Wang, J. Wang, Q. Du, and X. Yang, ‘‘Dog breed classification based on
deep learning,’’ in Proc. 13th Int. Symp. Comput. Intell. Design (ISCID),
Dec. 2020, pp. 209–212, doi: 10.1109/ISCID51228.2020.00053.
[2] X. Tu, K. Lai, and S. Yanushkevich, ‘‘Transfer learning on convolu-
tional neural networks for dog identification,’’ in Proc. IEEE 9th Int.
Conf. Softw. Eng. Service Sci. (ICSESS), Nov. 2018, pp. 357–360, doi:
10.1109/ICSESS.2018.8663718.
[3] Z. Ráduly, C. Sulyok, Z. Vadászi, and A. Zölde, ‘‘Dog breed identification
using deep learning,’’ in Proc. IEEE 16th Int. Symp. Intell. Syst. Informat.
(SISY), Sep. 2018, pp. 271–276, doi: 10.1109/sisy.2018.8524715.
[4] A. Varshney, A. Katiyar, A. K. Singh, and S. S. Chauhan, ‘‘Dog breed
images are taken as input which is converted to an array classification using deep learning,’’ in Proc. Int. Conf. Intell. Technol.
then it is passed to neural network. Train images are used (CONIT), Jun. 2021, pp. 1–5, doi: 10.1109/conit51480.2021.9498338.
to train our model. The training accuracy which we got is [5] R. O. Sinnott, F. Wu, and W. Chen, ‘‘A mobile application for dog breed
detection and recognition based on deep learning,’’ in Proc. IEEE/ACM
98.4% and the validation accuracy which we got is 92.4% 5th Int. Conf. Big Data Comput. Appl. Technol. (BDCAT), Dec. 2018,
for Hybrid(Inception-v3 +Xception). The training accuracy pp. 87–96, doi: 10.1109/BDCAT.2018.00019.
value for VGG19 algorithm which we got is 85% and the [6] M. V. S. Rishita and T. A. Harris, ‘‘Dog breed classifier using convolutional
neural networks,’’ in Proc. Int. Conf. Netw., Embedded Wireless Syst.
validation accuracy which we got is 55%. (ICNEWS), Dec. 2018, pp. 1–7, doi: 10.1109/ICNEWS.2018.8903980.
[7] A. Ayanzadeh and S. Vahidnia, ‘‘Modified deep neural networks
VI. CONCLUSION AND DISCUSSION for dog breeds identification,’’ 2018, Art. no. 2018120232, doi:
10.20944/preprints201812.0232.v1.
Considering the large number of breeds in this fine-grained [8] A. Sharma, A. Sahoo, M. Azhagiri, and D. Dutta, ‘‘Recognition of dog
classification problem, we view our overall findings as suc- breeds using convolutional neural network and visual geometry group,’’
cessful. Given the high variability both between and within Int. J. Eng. Adv. Technol., vol. 9, no. 1, pp. 3898–3902, Oct. 2019.
[9] B. V. Kumar and K. Bhavya, ‘‘Dog breed identification with fine tuning of
the 120 different breeds included in the dataset, we success- pre-trained models,’’ Int. J. Recent Technol. Eng. (IJRTE), vol. 8, nos. 2–
fully predict the right breed over 92.4% of the time in a 11, pp. 3677–3680, Sep. 2019.
single guess, a result that very few models could match. The [10] P. Borwarnginn, K. Thongkanchorn, S. Kanchanapreechakorn, and
W. Kusakunniran, ‘‘Breakthrough conventional based approach for dog
training accuracy which we got is 98.4% and the validation breed classification using CNN with transfer learning,’’ in Proc. 11th
accuracy which we got is 92.4% for Hybrid (Inception-v3 + Int. Conf. Inf. Technol. Electr. Eng. (ICITEE), Oct. 2019, pp. 1–5, doi:
Xception). The training accuracy value for VGG19 algorithm 10.1109/iciteed.2019.8929955.
[11] D. Kamdar, ‘‘Implementation of pre-trained deep learning model for dog
which we got is 85% and the validation accuracy which breed classification,’’ Turkish J. Comput. Math. Educ., vol. 120, no. 7,
we got is 55%. The proposed algorithms give a validation pp. 555–558, 2021.
accuracy score of 91.9% for Xception, 55% for VGG19, [12] D. D. D. Bhavani, M. H. S. Quadri, and Y. R. Reddy, ‘‘Dog breed iden-
tification using convolutional neural networks on Android,’’ CVR J. Sci.
83.47% for NASNetMobile, 89.05% for EfficientNetV2M, Technol., vol. 17, no. 1, pp. 62–66, Dec. 2019, doi: 10.32377/cvrjst1711.
87.38% for ResNet152V2, 92.4% for Hybrid of Inception- [13] S. M. Mohamed, ‘‘Dog breed identification using convolution neural
v3 & Xception, and 89.00% for Hybrid of EfficientNetV2M, network and web scraping,’’ Int. J. Sci. Res., vol. 9, no. 3, pp. 488–491,
Mar. 2020, doi: 10.21275/SR20307193747.
NASNetMobile, Inception & Xception. Among these algo- [14] K. Lai, X. Tu, and S. Yanushkevich, ‘‘Dog identification using soft biomet-
rithms, Hybrid of Inception-v3 & Xception gives the highest rics and neural networks,’’ in Proc. Int. Joint Conf. Neural Netw. (IJCNN),
accuracy of 92.4%. Jul. 2019, pp. 1–8, doi: 10.1109/IJCNN.2019.8851971.
[15] K. Mulligan and P. Rivas, ‘‘Dog breed identification with a neural network
Other machine learning models, neural networks, and deep over learned representations from the Xception CNN architecture,’’ in
learning models should all be further investigated in future Proc. 21st Int. Conf. Artif. Intell. (ICAI), 2019, pp. 1–6.

77238 VOLUME 11, 2023

B. Valarmathi et al.: Hybrid Deep Learning Algorithms for Dog Breed Identification

[16] D. M. Sundaram and A. Loganathan, ‘‘A new supervised clustering G. PRAKASH received the B.E. degree in com-
framework using multi discriminative parts and expectation–maximization puter science and engineering from Madras Uni-
approach for a fine-grained animal breed classification (SC-MPEM),’’ versity, in 1995, the M.E. degree in computer
Neural Process. Lett., vol. 52, no. 1, pp. 727–766, Aug. 2020. science and engineering from Annamalai Univer-
[17] G. Kanika, G. Sharankhala, A. Bhavya, and G. Riya, ‘‘Dog breed iden- sity, in 2004, and the Ph.D. degree in informa-
tification using CNN architecture,’’ Int. J. Sci. Eng. Res., vol. 11, no. 6, tion and communication engineering from Anna
pp. 1040–1045, 2020.
University, in 2015. He is currently an Associate
[18] I. Sneha, ‘‘Dog breed prediction using convolutional,’’ Neural Netw. Int. J.
Eng. Adv. Technol., vol. 9, no. 5, pp. 318–322, Jun. 2020.
Professor with the School of Computer Science
[19] T. Sadanandana, ‘‘Dog breed identification using pre-trained models,’’ and Engineering, Vellore Institute of Technology,
International Journal of Computer Science and Mobile Computing, vol. 9, Vellore. He has published 18 articles in Scopus and
no. 4, pp. 76–82, 2020. SCI-indexed journals and 16 papers in Springer and IEEE Xplore-sponsored
[20] Z. Wang, D. Zhao, and K. Hong, Dog Breed Classification. San Diego, CA, international conferences. His research interests include information secu-
USA: ECE Project Univ. California, 2020. rity, agile-based software engineering, cryptography, and steganography.
[21] P. Borwarnginn, W. Kusakunniran, S. Karnjanapreechakorn, and He is a Life Member of various professional bodies, such as ISTE, IAENG,
K. Thongkanchorn, ‘‘Knowing your dog breed: Identifying a dog breed and the Internet Society. He received various awards and certifications for
with deep learning,’’ Int. J. Autom. Comput., vol. 18, no. 1, pp. 45–54, his remarkable contributions to cyber security.
Feb. 2021, doi: 10.1007/s11633-020-1261-0.
[22] B. K. Shah, A. Kumar, and A. Kumar, ‘‘Dog breed classifier for
facial recognition using convolutional neural networks,’’ in Proc. 3rd
Int. Conf. Intell. Sustain. Syst. (ICISS), Dec. 2020, pp. 508–513, doi:
10.1109/ICISS49785.2020.9315871.
[23] D. Hsu, Using Convolutional Neural Networks to Classify Dog Breeds,
document CS231n, Convolutional Neural Networks for Visual Recogni- R. HEMADRI REDDY received the Ph.D. degree
tion, Feb. 2015. in mathematics, in 2007. He has ten years of teach-
[24] X. S. Higa, ‘‘Dog breed classification using convolutional neural
networks: Interpreted through a lockean perspective,’’ Senior
ing and research experience. He published more
Theses Student Publication, 2019. [Online]. Available: https://core. than 50 research articles in various reputed inter-
ac.uk/download/pdf/214315107.pdf national journals. His research interests include
[25] S. Kearney, W. Li, M. Parsons, K. I. Kim, and D. Cosker, ‘‘RGBD-dog: biofluid dynamics and machine learning.
Predicting canine pose from RGBD sensors,’’ in Proc. IEEE/CVF Conf.
Comput. Vis. Pattern Recognit. (CVPR), Jun. 2020, pp. 8333–8342, doi:
10.1109/cvpr42600.2020.00836.
[26] A. K. Agarwal, V. Kiran, R. K. Jindal, D. Chaudhary, and R. G. Tiwari,
‘‘Optimized transfer learning for dog breed classification,’’ Int. J. Intell.
Syst. Appl. Eng., vol. 10, no. 1s, pp. 18–22, Oct. 2022.
[27] N. H. Pratama, E. Rachmawati, and G. Kosala, ‘‘Classification of dog
breeds from sporting groups using convolutional neural network,’’ JIPI
(Jurnal Ilmiah Penelitian Dan Pembelajaran Informatika), vol. 7, no. 4,
pp. 1080–1087, Nov. 2022. S. SARAVANAN is currently an Assistant Pro-
[28] S. Ramaneswaran, K. Srinivasan, P. M. D. R. Vincent, and C.-Y. Chang, fessor with the Department of Electronics and
‘‘Hybrid inception v3 XGBoost model for acute lymphoblastic leukemia
Communication Engineering, Srinivasa Ramanu-
classification,’’ Comput. Math. Methods Med., vol. 2021, pp. 1–10,
jan Centre (SRC), SASTRA Deemed University,
Jul. 2021, doi: 10.1155/2021/2577375.
Kumbakonam, Tamil Nadu. Before his recent
B. VALARMATHI was born in Tirukovilur, Tamil appointment with SASTRA Deemed University,
Nadu, India. She received the degree in electron- he was an Associate Professor with CMIT, Ben-
ics and communication engineering and the mas- galuru. He has more than 20 years of experience
ter’s degree in computer science and engineering in both teaching and research. So far, he has pub-
from IIT Madras, India, and the Ph.D. degree in lished more than 70 Scopus and more than five
information and communication engineering from SCI/SCIE-indexed research articles in national and international journals.
Anna University, India. She has three decades of His research interests include VLSI design, hardware security, machine
teaching, research, and administrative experience. learning, and embedded systems. He is a Life Member of ISSE and ISTE.
She is currently a Professor of information tech-
nology with the Vellore Institute of Technology,
Vellore, India. She has published 47 research articles in data mining, machine
learning, sentiment analysis, natural language processing, text mining, the
Internet of Things, data science, soft computing, and heuristics. She is a Life
Member of the ISTE and the Soft Computing Research Society.
P. SHANMUGASUNDARAM received the Ph.D.
N. SRINIVASA GUPTA was born in Tiruvanna- degree from Anna University. He is currently an
malai, Tamil Nadu, India. He received the degree Associate Professor with the Department of Math-
in mechanical engineering and the master’s degree ematics, College of Natural and Computational
in industrial management from IIT Madras, India, Sciences, Mizan-Tepi University, Tepi Campus,
and the Ph.D. degree in mechanical engineering Ethiopia. His research title was ‘‘Applications of
from the Vellore Institute of Technology, Vellore, Intuitionist Fuzzy Sets in Decision Making Prob-
Tamil Nadu, India. He has three decades of teach- lems.’’ He has more than 26 years of experience in
ing, research, and administrative experience. He is teaching and more than 15 years of experience in
currently a Professor of mechanical engineering research. He has also been a peer reviewer of var-
with the Vellore Institute of Technology. He has ious Ph.D. theses, international journals, and conferences. He has published
published 25 research articles in cellular manufacturing, heuristics, data more than 25 Scopus/SCI-indexed research articles. His research interests
mining, sentiment analysis, natural language processing, the Internet of include fuzzy logic, machine learning, operations research, and research
Things, data science, soft computing, and text mining. He is a Life Member methodology. He is a life member of two professional bodies.
of ISTE.

VOLUME 11, 2023 77239