IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 14, No. 1, February 2025, pp. 482~491

ISSN: 2252-8938, DOI: 10.11591/ijai.v14.i1.pp482-491  482

Classification of Bharatanatyam postures using tailored

features and artificial neural network

Venkatesh Bhandage1, Basavaraj Anami2, Andrew J.1, Balachandra Hadimani3

1
Department of Computer Science and Engineering, Manipal Institute of Technology, Manipal Academy of Higher Education,
Manipal 576104, Karnataka, India
2
Department of Computer Science and Engineering, Faculty of Engineering, KLE Technological University, Hubballi, India
3
Department of Mathematics, Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal 576104, Karnataka,
India.

Article Info ABSTRACT

Article history: Bharatanatyam is a classical dance form of India that upholds the rich culture
of India. This dance is learned under the supervision of Guru, the teacher
Received Mar 2, 2024 traditionally called in India. The scarcity of experts resulted in the decline of
Revised Oct 16, 2024 people practicing this dance. There is a need for leveraging technology in
Accepted Oct 23, 2024 preserving and promoting this traditional dance and propagating it amongst
the youth. In this research, it is attempted to develop a methodology for
automated classification of Bharatanatyam dance postures. The methodology
Keywords: involves extraction of existing features such as speeded up robust features
(SURF) and histogram of oriented gradients (HOG), which are used to train
Adavu and test an artificial neural network (ANN). The results are corroborated with
AlexNet deep learning architectures such as AlexNet and GoogleNet. The proposed
Artificial neural network methodology has yielded a classification accuracy of 99.85% as compared
Bharatanatyam with 93.10% and 94.25% of AlexNet and GoogleNet respectively. The
GoogleNet proposed method finds applications such as assistance to Bharatanatyam
Histogram of oriented gradients dance teachers, e-learning of dance, and evaluating the correctness of the
Speeded up robust features postures.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Balachandra Hadimani
Department of Mathematics, Manipal Institute of Technology, Manipal Academy of Higher Education
Manipal 576104, Karnataka, India
E-mail: [email protected]

1. INTRODUCTION
Bharatanatyam is a prominent Indian classical dance (ICD) [1] with a rich cultural heritage [2], which
has embedded physical, mental, and emotional fitness qualities for the performers. Physical fitness is gained
with the movements of legs and hands for a given posture, mental health is attained by releasing stress with
happiness while performing during the concert, and emotional health is balanced by performing Navarasas,
nine forms of facial expressions [3], [4]. The cultural and historical foundations of Bharatanatyam are very
much evolved from the ancient temples of Tamil Nadu and South India. They use mudras, a form of
sophisticated hand gestures, to portray various devotional characters belonging to Indian mythology. Every
mudra has a specific meaning and significance. Along with the mudras, Bharatanatyam dance incorporates
postures, body movements, and facial expressions to convey a bunch of narratives and emotions. This intricate
combination of expressions and gestures brings the mythological stories to life, virtually carrying the viewers
into a world of spiritual depth and divine beauty. Bharatanatyam is the synchronized sequential combination
of different body postures, hand gestures, and facial expressions.

Journal homepage: http://ijai.iaescore.com

Int J Artif Intell ISSN: 2252-8938  483

To exhibit these mesmerizing performances requires years of dedicated learning by the performer.
Bharatanatyam students need to undergo intensive training to master the fundamental units such as mudras,
adavus, and navarasas. Adavus laid the foundation for body strength, grace, and coordination. As per the
scriptures, navarasas are the nine different human expressions. Abhinaya is an art of expression that helps the
dancers to convey a set of emotions to the audience. The interplay of hand gestures, body movements, postures,
footwork, and facial expressions allows the performer to become a storyteller, narrating the ancient Indian
mythological stories to the audience. Through its captivating performances and universal themes,
Bharatanatyam excels in cultural boundaries, promoting cross-cultural interactions and relations. With a sense
of identity among communities and feel of pride, Bharatanatyam stands as an exemplification of India's rich
cultural and artistic heritage.
This dance has to be learned from a Guru, a teacher, as it is governed by stringent and predefined rules
for performing different postures [5]. Nowadays, the youth is not fascinated much by this classical dance as
there are other modern dance forms across the world and a scarcity of experts to learn this dance. Since the youth
is more attracted to digital and portable devices, it is essential to take leverage of technological advancements,
such as digital image processing (DIP) [6]. Youth in today’s generation are more inclined towards digital gadgets
or mobile applications, and the essence of dance delivered through electronic equipment gains faster attention.
When Bharatanatyam dance is taught or delivered with a mobile application or web application, there are high
chances of gaining the attention and interest of the younger generation. Earlier researchers have worked on the
classification of human postures and dance poses by using shape-based features, such as, speeded up robust
features (SURF) and histogram of oriented gradients (HOG) with the help of machine learning classifiers, such
as artificial neural networks (ANN). There are very few researches carried out on the utilization of combinations
of features for pose classification, especially in the classification of Bharatanatyam postures. We have tried to
make use of a combination of shape-based features in our research. Since the Bharatanatyam dance has different
body postures combined with hand gestures, in this research we have presented a method for automated
classification of postures. The prominent body postures are shown in Figure 1.

Figure 1. Some of the postures enacted while performing Bharatanatyam

A total number of 26 postures are classified of which 18 postures are from two families of postures
called nattadavu and tattadavu and 8 postures are from namaskara postures, which are used for paying gratitude
(salutation) to Guru in India. We have given suitable names for the different namaskara postures, namely,
namaskara-1 to namaskara-8, which do not have exclusive names for individual postures. Similarly, nattadavu
postures are also given names from nattadavu-1 to nattadavu-10 for implementation reasons. Adavu is a
predefined sequence of postures [7]. For the purpose of research, we have considered the still images of those
postures. Some of the postures have pre-defined names associated with them. The list of postures considered
in the work is given in Table 1.

Table 1. List of Bharatanatyam postures considered

No. Posture name No. Posture name No. Posture name No. Posture name
1 Namaskara-1 8 Namaskara-8 15 Tattadavu-4 22 Nattadavu-6
2 Namaskara-2 9 Sthanaka 16 Aremandi-3 23 Nattadavu-7
3 Namaskara-3 10 Aremandi-1 17 Nattadavu-1 24 Nattadavu-8
4 Namaskara-4 11 Tattadavu-1 18 Nattadavu-2 25 Nattadavu-9
5 Namaskara-5 12 Tattadavu-2 19 Nattadavu-3 26 Nattadavu-10
6 Namaskara-6 13 Aremandi-2 20 Nattadavu-4
7 Namaskara-7 14 Tattadavu-3 21 Nattadavu-5

Objectives of the research carried out are i) to leverage technology to promote and propagate
Bharatanatyam dance; ii) to preserve the Indian cultural heritage and spiritual significance; and iii) to develop

Classification of Bharatanatyam postures using tailored features and artificial … (Venkatesh Bhandage)
484  ISSN: 2252-8938

a methodology for automated classification of Bharatanatyam dance postures. The contributions made in this
paper include i) classification of 26 still images of Bharatanatyam postures and classification from videos is
not considered within the scope of this paper; ii) adoption of hand-crafted features, namely, HOG and SURF
along with an ANN; and iii) pre-trained architectures such as AlexNet and GoogleNet are tested for the posture
dataset. The task is considered challenging because of the complexities, such as obstructions caused by the
attires of the performers, leading to misclassification of postures.
The remaining part of the article is organized into four sections. Section 2 gives a summary of the
related works in the literature. Section 3 describes the proposed methodology. A comparison with existing
works is given in section 4. Section 5 gives the conclusion.

2. RELATED WORK
We have carried out a literature survey to know about existing works in the classification of postures
and other related works connected to dance such as body pose recognition. The idea of cited papers is as under.
Machine learning and deep learning models are used in [3], [5] for the classification of ICD. The widely used
deep learning architecture, AlexNet, is introduced by [8] and is adopted in research [9] for the recognition of
human beings. The 22-layer deeper GoogleNet architecture is introduced by Szegedy et al. [10]. Multiple
classifiers for the Indian sign language (ISL) classification are adopted in [11]. The most widely used scale and
rotation invariant shape descriptors, SURF are introduced by Bay et al. [12]. Combinations of hand-crafted
features and convolutional neural network (CNN) features are utilized in [13] for the classification of
single-hand gestures of ICD. These hybrid feature vectors have given a classification accuracy of 95% when
experimented with the VGG16 model. It is evident that hybrid features can play a better role in the classification
of mudras. Attention-based features are used by researchers in [14] for recognition of Bharatanatyam poses.
Deep learning architectures are effectively used in [15] for the identification of human beings. Ensembling of
classifiers [16] and key points [17] are utilized in hand gesture and pose classification, respectively.
Dalal and Triggs [18] have presented a HOG features for detecting human beings. Deep learning has also been
proven to be effective in classifying fruits [19] along with their ability to classify ICD poses [20].
The HOG features are found to be relevant in estimating the orientation of the human body [21]. Deep
learning is used by [22] for human pose estimation. Researchers have worked on dance pose recognition [23],
dance choreography [24], and annotation of dance frames [25]. A combination of features, local self-similarity
(HOG-LSS) is adopted in [26] for pedestrian detection and is found to be useful. Researchers in [27]–[32] have
used shape-based features with machine learning classifiers and adopted deep learning architectures in
classifying hand gestures, mudras, and postures of ICD. However, they have not attempted a combination of
features. A rule-based approach utilizing key points for classifying Bharatanatyam mudras has given an
accuracy of 72.04% [33]. Machine learning and deep learning classifiers are adopted in the detection of yoga
poses [34], recognition of sitting postures [35], recognition of Bharatanatyam mudras [36], and recognition of
dance movement [37].
In summary, certain works are attempted in recognizing human poses, classifying ICD and adavus,
and identifying human beings using pre-trained deep learning architectures. Appearance-based features are
used. There is less research carried out on applying a combination of shape-based features for the classification
of Bharatanatyam poses. Deep learning approaches are adopted for the classification of human body poses and
ICD poses. Since postures are fundamental units of Bharatanatyam, a work on the classification of different
poses of postures is carried out.

3. PROPOSED METHODOLOGY
The methodology involves three stages, namely data acquisition, obtaining features, and posture
classification, as shown in Figure 2. The setup used for data acquisition is given in Figure 3. A uniform black
background is fixed before the acquisition of images.

Figure 2. Proposed methodology for classification of postures

Int J Artif Intell, Vol. 14, No. 1, February 2025: 482-491

Int J Artif Intell ISSN: 2252-8938  485

Figure 3. Process of capturing posture images

3.1. Feature extraction

The SURF and HOG features are extracted from captured posture images and deployed in
classification. The dataset images are converted to grayscale before extracting the features. Figure 4 shows the
process of extracting SURF features. The posture of nattadavu-9 and its grayscale image are depicted in
Figures 4(a) and 4(b), respectively. The number of key points chosen is 5 and 64 features are extracted
concerning each of the key points, as shown in Figure 4(c). A total of 320 feature values are obtained from
each of the postures and used to train and test ANN.

(a) (b) (c)

Figure 4. Scenario of obtaining SURF features from Nattadavu-9 posture image of (a) Nattadavu-9
posture, (b) gray form of posture, and (c) extraction of SURF features

The HOG features are extracted by segmenting out the posture part from the background as depicted
in Figure 5. Nattadavu-9 posture, extracted region of interest (ROI) of posture, the gray form of the ROI image,
and the image resized to 128×64 are given in Figures 5(a) to 5(d) respectively. A bounding box is used to
extract the posture part from the image, which is converted to grayscale and later resized to 128×64 pixels, the
size used for calculation of HOG features [18]. A cell size of 16×16 is used for calculating HOG features and
a total of 756 features are obtained for each posture image. The sample SURF feature values out of 320 features
and sample HOG features out of 756 features, extracted from the nattadavu-9 posture image, are given in
Table 2. The time elapsed in obtaining the adapted features, on groups of individual posture images, is also
given in Table 2. The total time required for calculation of SURF features of our image dataset is 94.39 seconds
and the time required for calculation of HOG features is 39.66 seconds. The experiments are conducted on a
core i3-8100 processor at 3.60 GHz with a memory of 8 GB.

Classification of Bharatanatyam postures using tailored features and artificial … (Venkatesh Bhandage)
486  ISSN: 2252-8938

(a) (b) (c) (d)

Figure 5. Scenario of obtaining HOG features from nattadavu-9 posture image of (a) nattadavu-9 posture,
(b) ROI of posture is extracted, (c) gray form of the ROI image, and (d) image resized to 128×64

Table 2. SURF and HOG features of nattadavu-9 posture and time estimates of all the postures
SURF features HOG features Time estimate of SURF and HOG features
No. Value No. Value Posture No. SURF HOG Posture No. SURF HOG
1 0.000135 1 0.141226 1 5.46 1.23 14 3.04 1.56
2 4.67E-05 2 2.64E-01 2 4.78 1.68 15 2.99 1.61
3 0.000338 3 0.286343 3 7.65 1.85 16 3.20 1.60
4 8.01E-05 4 1.25E-01 4 3.05 1.57 17 3.09 1.62
. . . . 5 5.79 1.46 18 3.18 1.66
. . . . 6 4.00 1.58 19 3.12 1.52
7 3.47 1.63 20 3.23 1.67
8 3.37 1.42 21 3.05 1.53
9 3.30 1.19 22 3.05 1.40
317 0.028662 753 0.124955 10 3.39 1.39 23 3.04 1.49
318 0.000175 754 0.030808 11 3.14 1.31 24 3.36 1.62
319 0.00017 755 0.108684 12 3.19 1.52 25 3.13 1.35
320 0.0002 756 0.382456 13 3.06 1.73 26 3.26 1.47

3.2. Classification
The dataset consisting of images of postures is created in natural environment. Both male and female
dancers of different age groups are considered. MATLAB 2018a is utilized to implement the methodology.
The dataset has a total of 6500 images, which include the images captured from both male and female
Bharatanatyam dancers of different age groups, more than 10 dancers in each group. The devised methodology
for the classification of postures is depicted in Figure 2. ANN is trained and tested for posture classification by
using HOG and SURF features. The data augmentation is not performed during the process of model training
as there are postures like nattadavu-5 and nattadavu-9 are looking flipped but they are different postures, as
they are performed from two sides of the human body. Hence, the data augmentation techniques like flipping,
and rotating are not done during the training process. The combined feature set SURF-HOG is also tested to
classify postures. The ANN is constructed by using the neural network (NN) toolbox of MATLAB 2018a.
Table 3 gives the parameters used to set up a neural network.

Table 3. Hyperparameters of ANN

Parameter Used value Parameter Used value
Number of hidden layers 1 Transfer function Log sigmoid
Metric used for Accuracy Number of neurons used at output layer 26
Function for learning Learngdm Number of neurons used at hidden layer 640 | 1512 | 2152
Function for learning Traingdx Number of neurons used at input layer 320 | 756 | 1076
Type of network adopted Backpropagation Initial learning rate 0.0001

When experimented with SURF features, a classification accuracy of 98.55% is obtained. Figure 6
shows some of the postures that conflict. For example, namaskara-1 and namaskara-4 postures are conflicting.
Sthanaka and aremandi-1 postures are conflicting. Tattadavu-3, tattadavu-4, and aremandi-3 are conflicting.
Nattadavu-5 and nattadavu-9 are conflicting. When experimented with HOG features, a classification accuracy
of 98.71% is obtained. Postures Namaskara-1 and namaskara-2, namaskara-1 and sthanaka, aremandi-2 and
aremandi-3 are found conflicting. When experimented with combined SURF-HOG features, a classification
accuracy of 99.85% is obtained. The classification results produced by ANN are depicted in Figure 7.

Int J Artif Intell, Vol. 14, No. 1, February 2025: 482-491

Int J Artif Intell ISSN: 2252-8938  487

Namaskara-1 Namaskara-2 Namaskara-4 Sthanaka Nattadavu-5 Nattadavu-9

Aremandi-1 Tattadavu-3 Tattadavu-4 Aremandi-2 Aremandi-3

Figure 6. Some of the conflicting postures

Figure 7. Classification results of each posture obtained by using SURF+HOG features

The ANN classification results are measured in terms of precision, recall, accuracy, and F1 score.
These classification measures are given in (1) to (4) in which true positive (TP), false positive (FP), true
negative (TN), and false negative (FN), respectively. The micro-averaging technique is adopted to calculate
these classification measures. In the case of micro averaging, sums of FPs, FNs, and TPs are found to compute
the global average F1 score. The obtained classification results are given in Table 4.

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = 𝑇𝑃/(𝑇𝑃 + 𝐹𝑃) (1)

𝑅𝑒𝑐𝑎𝑙𝑙 = 𝑇𝑃/(𝑇𝑃 + 𝐹𝑁) (2)

𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = (𝑇𝑃 + 𝑇𝑁)/(𝑇𝑃 + 𝑇𝑁 + 𝐹𝑃 + 𝐹𝑁) (3)

𝐹1 𝑠𝑐𝑜𝑟𝑒 = 2𝑇𝑃/(2𝑇𝑃 + 𝐹𝑃 + 𝐹𝑁) (4)

Table 4. Classification results of ANN with considered features

Precision (%) Recall (%) Accuracy (%) F1-Score (%) Iterations Training time
SURF 94.17 94.17 98.55 94.17 1339 95
HOG 96.27 96.27 98.71 96.27 751 318
SURF-HOG 98.13 98.13 99.85 98.13 935 732

To corroborate the results obtained by hand-crafted features deep learning architectures are tested.
Deep learning applications significantly use the process of transfer learning. Two pre-trained deep learning
architectures, namely, AlexNet and GoogleNet are used.
The AlexNet is trained with millions of images and is capable of classifying 1000 different object
classes, such as keyboards, pencils, and animals. A fine-tuning of an AlexNet is made to suit the ICD posture
classification. The architecture is made up of three fully connected layers and five convolutional layers. The

Classification of Bharatanatyam postures using tailored features and artificial … (Venkatesh Bhandage)
488  ISSN: 2252-8938

input image of size 224×224×3 is fed to 1st convolution layer having 96 11×11×3 sized kernels, with a
4-pixel stride. The first layer output is normalized, pooled, and given as an input for the second layer. The
filters with 256 kernels of size 5×5×48 is used at the second layer. No normalization and pooling layers are
present between the 5th, 4th, and 3rd convolutional layers. There are 384 kernels, of size 3×3×256, in 3rd layer.
There are 384 kernels, of size 3×3×192, in the 4th layer and 256 kernels with size 3×3×192 in the 5th layer [2].
The last 3 layers of this pre-trained architecture are fine-tuned for the Bharatanatyam posture image
classification problem. The 3 layers are replaced with a fully connected layer, a softmax layer, and a
classification output layer. As there are 26 classes in our problem, the size of the fully connected layer is set to
26. The images in the dataset are normalized to 227×227×3 as per the requirements of the architecture. The
posture dataset is appropriately divided into a train set and a test set. The learning rate is set to 0.0001 and the
setup has involved a single CPU, with 8 GB memory. The training cycle involved 1 epoch with 114 iterations.
The total time elapsed to train and test our network is 9 minutes and 47 seconds. A validation accuracy of
93.10% is obtained.
GoogleNet, another pre-trained architecture, is tested on the posture dataset, wherein the images are
normalized to 224×224×3, as per the requirements of the architecture. The GoogleNet, with increased width
and depth, is a 22-layer deep network. It has inception, max pooling, and convolutional layers, which are core
components of this architecture. The inception layer is a combination of 1×1, 3×3, and 5×5 convolutional
layers. The inception layer output is concatenated as a single vector and is provided as input for the next stage
[10]. The learning rate is set to 0.0001 and the setup involves a single CPU. The training cycle involved 6
epochs with 182 iterations per epoch, resulting in a total of 1092 iterations. The total time elapsed for training
and testing the GoogleNet is set to 254 minutes and 37 seconds. GoogleNet has resulted in a validation accuracy
of 94.25%. The summary of ANN and deep learning-based classifications is given in Figure 8.

Figure 8. Classification results of Alexnet, Googlenet, and ANN classifiers with SURF, HOG, and
SURF+HOG features

4. DISCUSSION AND COMPARATIVE STUDY

The discussion involves the interpretation of the obtained results from an ANN classifier and the
comparison of considered SURF and HOG features in terms of their classification results and time efficiency.
The SURF feature has produced a classification accuracy of 98.55%. When experimented with SURF features,
a total of 4 postures are classified with an accuracy of 100%, which amounts to 15.38% of the total postures
considered in the work. The remaining 24 postures are classified with accuracies ranging above 98% and less
than 100%. The postures in the dataset are classified with accuracies not less than 98%. Some of the postures
conflict with others in the dataset, which resulted in misclassification or reduced classification accuracy. This
shows that SURF features are suitable to classify Bharatanatyam postures. The HOG feature has produced a
classification accuracy of 98.71%. When experimented with HOG features, postures namaskara-1 and
namaskara-2, namaskara-1 and sthanaka, aremandi-2 and aremandi-3 are found conflicting with each other. A
total of 12 postures are classified with an accuracy of 100%, which amounts to 46.15% of the total postures
considered in the work. The remaining 14 postures are classified with accuracies ranging above 98% and less
than 100%. The postures in the dataset are classified with accuracies not less than 98%. This shows that HOG
features can also better classify Bharatanatyam postures. The combined feature, SURF-HOG has produced an
improved classification accuracy of 99.85%. When experimented with combined SURF-HOG features, a total
of 14 postures are classified with an accuracy of 100%, which amounts to 50% images in the considered dataset.
The remaining 14 postures are classified with accuracies greater than 99%. The postures in the dataset are
classified with accuracies not less than 99%. With the results, it is evident that the misclassification of
conflicting postures is reduced and overall classification accuracy has improved with the combined
SURF-HOG feature. This classification accuracy is better compared to deep learning architectures, AlexNet
and GoogleNet, which have given classification accuracies of 93.10 and 94.25%, respectively.

Int J Artif Intell, Vol. 14, No. 1, February 2025: 482-491

Int J Artif Intell ISSN: 2252-8938  489

The time elapsed for training and testing ANN with SURF features is 1 minute and 35 seconds, with
HOG features is 5 minutes and 18 seconds and with combined SURF-HOG feature is 12 minutes and 12
seconds. This shows that the time required by ANN for classifying postures with combined SURF-HOG
features is more compared to individual SURF and HOG features. Even though the time required is more for
classification with combined SURF-HOG features, the accuracy obtained for classification is more compared
to individual features. The time elapsed for training AlexNet with the considered dataset is 9 minutes and
47 seconds. The time taken for training GoogleNet is 254 minutes and 37 seconds. The time taken by deep
learning classifiers is more compared to ANN. The classification accuracies produced by AlexNet and
GoogleNet, on our dataset, are less compared to the accuracy obtained by ANN with the combined
SURF-HOG feature. With the results, it is evident that the misclassification of conflicting postures is reduced
and overall classification accuracy has improved with the combined SURF-HOG feature. With these
experiments, it is evident that the combined SURF-HOG feature is effective in terms of classification measures
and time required for classification when compared with AlexNet and GoogleNet classifiers.

4.1. Comparative study

The proposed methodology is compared with existing and related works as presented in Table 5. The
work is compared in terms of the problem domain, datasets, and classifiers used. The obtained classification
accuracies are promising. Earlier researchers have used machine learning and deep learning classifiers for the
classification of dance poses. Some researchers have used existing datasets and some researchers have created
their own dataset. The number of classes and the corresponding reported classification results are less compared
to our research. There are few attempts made in the utilization of a combination of features. Compared to earlier
researchers, our research has attempted to use combination of shape-based features and more number of classes.
Our methodology has produced better classification accuracy compared to earlier research.
Our research finds applications in developing mobile applications or web-based applications for
assisting novice learners. Interactive applications can be developed to guide young learners and help them to
correct their poses, in case of deviations by the learners. Our research has scope for further study. We have
considered only 26 poses, containing poses of a few adavus and namaskara. There is scope for considering the
classification of other adavus and the classification of facial expressions of Bharatanatyam dancers. Further
research can be carried out in the direction of classification of Bharatanatyam postures from videos.
Explainable artificial intelligence could be adapted in further study.

Table 5. Comparison with state-of-the-art works cited in the literature

Number Classification
Ref Dataset Technique Experimental data
of classes accuracy (%)
[9] ICD Mudra Dataset SURF and HOG features are used 120 images of poses 24 98
with SVM
[13] Asamyukta Mudra dataset Hu-Moments and VGG features 2610 images with 90 29 95
images of each class
[19] Bharatanatyam pose Genetic algorithm for pose generation Skeletal of images 25 80
[27] Bharatnatyam, Kathak, SVM classifier 100 images 15 90
and Odissi
[32] Dance pose Body skeleton is utilized with LSTM Skeleton data of poses 10 95.2
[33] Hand gesture dataset Rules devised using Euclidean Keypoint images of 31 72.04
distance between joint angles MediaPipe
[37] Balletto dance Pose estimation algorithm and Skeletal nodes of the 10 95.2
LSTM human body
Our work Bharatanatyam Adavu SURF + HOG with ANN 6500 images of Adavu 26 99.85

5. CONCLUSION
The classification of 26 classes of Bharatanatyam postures is attempted in this work. The handcrafted
features SURF, HOG, and combined SURF-HOG features are adopted and the classification results are
measured in terms of precision, recall, accuracy, and F1 score. The time required for the calculation of
considered tailored features is analyzed. ANN is used for classification using hand-crafted features. The SURF,
HOG, and combined SURF-HOG features have produced classification accuracies of 98.55, 98.71, and
99.85%, respectively, on the considered dataset. To corroborate the obtained results, the pre-trained deep
learning architectures, namely, AlexNet and GoogleNet are adopted and fine-tuned for the classification task.
The validation accuracies of 93.10 and 94.25% are produced by AlexNet and GoogleNet, respectively. The
time taken by ANN, AlexNet, and GoogleNet classifiers, for classification of images in our posture dataset,
are compared. The combined SURF-HOG has produced better classification results over the deep learning
architectures. Compared to existing works, the present work has used a larger number of images, and the results

Classification of Bharatanatyam postures using tailored features and artificial … (Venkatesh Bhandage)
490  ISSN: 2252-8938

are corroborated with deep learning architectures. The proposed methodology can be adopted for the
classification of other postures of Bharatanatyam and other human body postures, in general. Some of the
envisaged applications of the method include evaluation of postures performed by Bharatanatyam dancers,
e-learning of this ICD, and possibly delivery of automated commentary at the time of concerts.

REFERENCES
[1] S. Jadhav and J. D. Pawar, “Aesthetics of bharatanatyam poses evaluated through fractal analysis,” in Advances in Intelligent
Systems and Computing, Springer, Singapore, 2017, pp. 401–409, doi: 10.1007/978-981-10-2471-9_39.
[2] K. V. V. Kumar and P. V. V. Kishore, “Indian classical dance mudra classification using HOG features and SVM classifier,” in
Smart Innovation, Systems and Technologies, Springer, Singapore, 2018, pp. 659–668, doi: 10.1007/978-981-10-5544-7_65.
[3] A. Mohanty et al., “Nrityabodha: towards understanding indian classical dance using a deep learning approach,” Signal Processing:
Image Communication, vol. 47, pp. 529–548, Sep. 2016, doi: 10.1016/j.image.2016.05.019.
[4] A. Mohanty and R. R. Sahay, “Rasabodha: understanding Indian classical dance by recognizing emotions using deep learning,”
Pattern Recognition, vol. 79, pp. 97–113, Jul. 2018, doi: 10.1016/j.patcog.2018.01.035.
[5] K. V. V. Kumar, P. V. V. Kishore, and D. A. Kumar, “Indian classical dance classification with adaboost multiclass classifier on
multifeature fusion,” Mathematical Problems in Engineering, vol. 2017, no. 1, Jan. 2017, doi: 10.1155/2017/6204742.
[6] R. C. Gonzalez and R. E. Woods, Digital image processing, 3rd ed. Upper Saddle River, New Jersey: Pearson Education
International, 2009.
[7] A. Mohanty, K. Roy, and R. R. Sahay, “Nrityamanthan: unravelling the intent of the dancer using deep learning,” in Heritage
Preservation, Singapore: Springer Singapore, 2018, pp. 209–239, doi: 10.1007/978-981-10-7221-5_11.
[8] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet classification with deep convolutional neural networks,” Communications
of the ACM, vol. 60, no. 6, pp. 84–90, May 2017, doi: 10.1145/3065386.
[9] A. A. Almisreb, N. Jamil, and N. M. Din, “Utilizing alexnet deep transfer learning for ear recognition,” in 2018 Fourth International
Conference on Information Retrieval and Knowledge Management (CAMP), Mar. 2018, pp. 1–5, doi: 10.1109/INFRKM.2018.8464769.
[10] C. Szegedy et al., “Going deeper with convolutions,” in 2015 IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), IEEE, Jun. 2015, pp. 1–9, doi: 10.1109/CVPR.2015.7298594.
[11] H. Bhavsar and D. J. Trivedi, “Performance comparison of SVM, CNN, HMM and neuro-fuzzy approach for Indian sign language
recognition,” Indian Journal of Computer Science and Engineering (IJCSE), vol. 12, no. 4, pp. 1093–1101, Aug. 2021, doi:
10.21817/indjcse/2021/v12i4/211204220.
[12] H. Bay, A. Ess, T. Tuytelaars, and L. V. Gool, “Speeded-up robust features (SURF),” Computer Vision and Image Understanding,
vol. 110, no. 3, pp. 346–359, Jun. 2008, doi: 10.1016/j.cviu.2007.09.014.
[13] R. Santhosh and R. Kk, “Classification of asmyukta mudras in Indian classical dance using hadncrafted and pre-trained features
with machine learning and deep learning methods,” SSRN, pp. 1-22, 2024, doi: 10.2139/ssrn.4818826.
[14] D. A. Kumar, P. V. V. Kishore, and K. Sravani, “Deep bharatanatyam pose recognition: a wavelet multi head progressive attention,”
Pattern Analysis and Applications, vol. 27, no. 2, Jun. 2024, doi: 10.1007/s10044-024-01273-0.
[15] M. M. A. Abuqadumah, M. A. M. Ali, A. A. Almisreb, and B. Durakovic, “Deep transfer learning for human identification based
on footprint: a comparative study,” Periodicals of Engineering and Natural Sciences (PEN), vol. 7, no. 3, Sep. 2019, doi:
10.21533/pen.v7i3.733.
[16] R. V. Amrutha and G. Malu, “EnGesto: an ensemble learning approach for classification of hand gestures,” IEEE Access, vol. 12,
pp. 85709–85723, 2024, doi: 10.1109/ACCESS.2024.3411155.
[17] Z. Jiang, H. Ji, C.-Y. Yang, and J.-N. Hwang, “2D human pose estimation calibration and keypoint visibility classification,” in
ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE, Apr. 2024, pp.
6095–6099, doi: 10.1109/ICASSP48485.2024.10448474.
[18] N. Dalal and B. Triggs, “Histograms of oriented gradients for human detection,” in 2005 IEEE Computer Society Conference on
Computer Vision and Pattern Recognition (CVPR’05), IEEE, pp. 886–893, doi: 10.1109/CVPR.2005.177.
[19] N. A. Muhammad, A. Ab Nasir, Z. Ibrahim, and N. Sabri, “Evaluation of CNN, alexnet and googlenet for fruit recognition,” Indonesian
Journal of Electrical Engineering and Computer Science, vol. 12, no. 2, pp. 468-475, Nov. 2018, doi: 10.11591/ijeecs.v12.i2.pp468-475.
[20] P. V. V. Kishore et al., “Indian classical dance action identification and classification with convolutional neural networks,”
Advances in Multimedia, vol. 2018, pp. 1–10, 2018, doi: 10.1155/2018/5141402.
[21] K. Panachit, O. S. Guat, and E. How-Lung, “Estimation of human body orientation using histogram of oriented gradients,” in
MVA2011 IAPR Conference on Machine Vision Applications, Nara, JAPAN, pp. 459-462, 2011.
[22] P. Szczuko, “Deep neural networks for human pose estimation from a very low resolution depth image,” Multimedia Tools and
Applications, vol. 78, no. 20, pp. 29357–29377, Oct. 2019, doi: 10.1007/s11042-019-7433-7.
[23] S. Saha, A. Banerjee, S. Basu, A. Konar, and A. K. Nagar, “Fuzzy image matching for posture recognition in ballet dance,” in 2013
IEEE International Conference on Fuzzy Systems (FUZZ-IEEE), Jul. 2013, pp. 1–8, doi: 10.1109/FUZZ-IEEE.2013.6622401.
[24] S. Jadhav, M. Joshi, and J. Pawar, “Art to smart: an automated bharatanatyam dance choreography,” Applied Artificial Intelligence,
vol. 29, no. 2, pp. 148–163, Feb. 2015, doi: 10.1080/08839514.2015.993557.
[25] S. S. and D. J. M V, “Automatic annotation of dance videos based on foot postures,” Indian Journal of Computer Science and
Engineering (IJCSE), vol. 11, no. 1, pp. 89–98, Feb. 2020, doi: 10.21817/indjcse/2020/v11i1/201101047.
[26] S. Yao, S. Pan, T. Wang, C. Zheng, W. Shen, and Y. Chong, “A new pedestrian detection method based on combined HOG and
LSS features,” Neurocomputing, vol. 151, pp. 1006–1014, Mar. 2015, doi: 10.1016/j.neucom.2014.08.080.
[27] Shubhangi and U. S. Tiwary, “Classification of Indian classical dance forms,” in Intelligent Human Computer Interaction, Springer,
Cham, 2017, pp. 67–80, doi: 10.1007/978-3-319-52503-7_6.
[28] S. Saha, S. Ghosh, A. Konar, and A. K. Nagar, “Gesture recognition from indian classical dance using kinect sensor,” in 2013 Fifth
International Conference on Computational Intelligence, Communication Systems and Networks, IEEE, Jun. 2013, pp. 3–8, doi:
10.1109/CICSYN.2013.11.
[29] R. Jisha Raj, S. Dharan, and T. T. Sunil, “Classification of indian classical dance hand gestures: a dense sift based approach,” in
Algorithms for Intelligent Systems, Springer, Singapore, 2022, pp. 181–195, doi: 10.1007/978-981-16-6893-7_17.
[30] B. S. Anami and V. A. Bhandage, “SME feature-based classification of adavu and posture images of bharatanatyam dance,”
International Journal of Arts and Technology (IJART), vol. 12, no. 4, 2020, doi: 10.1504/IJART.2020.112640.
[31] B. S. Anami and V. A. Bhandage, “Suitability study of certain features and classifiers for bharatanatyam double-hand mudra

Int J Artif Intell, Vol. 14, No. 1, February 2025: 482-491

Int J Artif Intell ISSN: 2252-8938  491

images,” International Journal of Arts and Technology (IJART), vol. 11, no. 4, 2019, doi: 10.1504/IJART.2019.103497.
[32] B. S. Anami and V. A. Bhandage, “A comparative study of suitability of certain features in classification of bharatanatyam mudra images
using artificial neural network,” Neural Processing Letters, vol. 50, no. 1, pp. 741–769, Aug. 2019, doi: 10.1007/s11063-018-9921-6.
[33] S. Paul and P. P. Das, “Fast detection and rule based classification of bharatanatyam hasta mudra,” in Communications in Computer
and Information Science, Springer, Cham, 2023, pp. 589–603, doi: 10.1007/978-3-031-31407-0_44.
[34] A. K. Rajendran and S. C. Sethuraman, “A survey on yogic posture recognition,” IEEE Access, vol. 11, pp. 11183–11223, 2023,
doi: 10.1109/ACCESS.2023.3240769.
[35] J. Fang, S. Shi, Y. Fang, and Z. Huo, “Realtime sitting posture recognition on embedded device,” in MultiMedia Modeling, Springe,
Cham, 2023, pp. 313–324, doi: 10.1007/978-3-031-27818-1_26.
[36] N. Parthasarathy and Y. Palanichamy, “Novel video benchmark dataset generation and real-time recognition of symbolic hand
gestures in indian dance applying deep learning techniques,” Journal on Computing and Cultural Heritage, vol. 16, no. 3,
pp. 1–19, Sep. 2023, doi: 10.1145/3586207.
[37] P. Lei, N. LI, and H. Liu, “Dance movement recognition based on gesture,” in Proceedings of the 2nd International Conference on
Internet, Education and Information Technology (IEIT 2022), 2023, pp. 448–452, doi: 10.2991/978-94-6463-058-9_73.

BIOGRAPHIES OF AUTHORS

Venkatesh Bhandage is presently working as Assistant Professor-Senior in the

Department of CSE Manipal Institute of Technology, MAHE, Manipal. He worked in the
Department of CSE at TCE, Gadag, for more than 10 years. He also worked in the Department
of ISE at SKSVMACET, Laxmeshwar, Gadag for 1 and a half years. He worked as an intern at
VMware, Bangalore for one year. He has 15 years of professional experience and 8 years of
research experience. He received his Ph.D. from Visvesvaraya Technological University,
Belagavi in 2020. He pursued his M.Tech. in Computer Engineering from SJCE, Mysore in
2009 and B.E. in CSE from Basaveshwara Engineering College, Bagalkot in 2007. His research
interests include image processing, artificial intelligence, medical image processing, and
machine learning. He has published 14 research articles in reputed journals and conferences. He
is a member of IEEE, IE, and ISTE. He can be contacted at email:
[email protected] or [email protected].

Basavaraj Anami has pursued his Ph.D. from University of Mysore, Mysore in the
year 2003. He has obtained his M.Tech. (Computer Science), from IIT Madras in the year 1986.
He has obtained his Engineering degree from Karnataka University Dharawad in the year 1981.
He has 40+ years of professional experience and 20+ years of research experience. His research
interests include image processing, artificial intelligence and machine learning. He has
published 110+ papers in international journals nad conferences. He has worked as Principal,
K. L. E. Institute of Tchnology, Hubballi for more than 16 years. He is presently working as
Registrar at K. L. E. Tchnological University, Hubballi. He can be contacted at email:
[email protected].

Andrew J. is currently working as an Associate Professor in the Department of

Computer Science and Engineering (CSE) at Manipal Institute of Technology (MIT), Manipal,
India. He received his Ph.D. from the Vellore Institute of Technology (VIT), Vellore, India. He
received his Bachelor of Engineering (B.E.) in CSE in 2011 and received his Master of
Engineering (M.E.) degree in 2013 from Anna University, Chennai, India. He is an active
researcher who has published more than 50 scientific research articles in reputed journals and
conferences. He also served as a speaker at prestigious conferences worldwide. He has 10 years
of teaching experience. His research interest includes data privacy, healthcare data analysis,
deep learning, machine learning, computer vision, and blockchain technologies. He can be
contacted at email: [email protected].

Balachandra Hadimani is presently working as Assistant Professor in the

Department of Mathematics Manipal Institute of Technology, MAHE, Manipal. He has 14 years
and professional experience and 10 years of research experience. He has pursued M.Sc. in
Mathematics and Ph.D. in Differential Geometry. His research interests include differential
geometry and biomechanics. He has published 13 papers in international journals. He can be
contacted at email: [email protected].

Classification of Bharatanatyam postures using tailored features and artificial … (Venkatesh Bhandage)