Neural Computing and Applications (2022) 34:17783–17792

https://doi.org/10.1007/s00521-022-07404-0 (0123456789().,-volV)(0123456789().
,- volV)

ORIGINAL ARTICLE

High embedding capacity in 3D model using intelligent Fuzzy based

clustering
Modigari Narendra1 • M. L. Valarmathi2 • L. Jani Anbarasi1 • M. Vergin Raja Sarobin1 • Fadi Al-Turjman3

Received: 14 February 2021 / Accepted: 6 May 2022 / Published online: 4 June 2022
Ó The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2022

Abstract
High embedding capacity plays a vital role in the watermarking schemes to secure the ownership, authenticity, and
copyright-related issues of 3D models. This study describes the watermark embedding using the High Embedding Capacity
in 3D Models based on Intelligent Fuzzy Clustering scheme that includes standard Fuzzy C-means (HE-FCM) and
Intelligent Fuzzy C-Means (HE-IFCM). Efficiency of optimized cluster segmentation is evaluated and compared with other
optimization algorithms. Bit embedding intervals are identified in the resultant clusters to embed the watermark data
resulting high embedding capacity. The embedded watermark can be extracted by performing the reverse operation. The
embedding capacity, distortion rate and robustness are analyzed using Peak Signal to Noise Ratio (PSNR), Root Means
Square Error, correlation coefficient and Hausdorff distance. This scheme is capable of embedding 3 KB to 15 KB for
small models and 164KB to 820 KB for larger models achieving around 68 dB of PSNR with less distortion. HE-IFCM is
successful in embedding large capacity of watermark data resulting in minimal distortion.

Keywords 3D model  Embedding capacity  FCM  Intelligent FCM

1 Introduction graphics and the animation process. Due to the close link
among the model skeleton and the boundaries, the seg-
Mesh segmentation of shapes is a well-researched area mentation and the skeletonization becomes an important
with challenging problems yet to be solved for 3D mesh issue. Some models remain consistent in all poses whereas
models. It is the fundamental concepts in computer few other models are inconsistent. Consistency among all
the families of objects has to be achieved using any seg-
mentation algorithm. Common mesh segmentation usually
& Modigari Narendra does not perform any deformation in the object geometry
[email protected]
directly. Instead mostly it tries to identify the boundaries
M. L. Valarmathi between the object parts for segmentation. Most of the
[email protected]
algorithms sets an objective functions before the segmen-
L. Jani Anbarasi tation process, if not they try to search for the concavity of
[email protected]
the surface models.
M. Vergin Raja Sarobin The part based segmentation includes volume based,
[email protected]
surface based or skeleton based algorithms. Volume based
Fadi Al-Turjman segmentation includes approximate convex decomposition,
[email protected]
volumetric meshes or space partitioning. Similarly surface
1
School of Computer Science and Engineering, Vellore based segmentation includes region growing, watershed
Institute of Technology, Chennai, India algorithm, iterative or hierarchical clustering and boundary
2
Department of Computer Science and Engineering, Dr. algorithms for segmenting the 3D models. Skeleton based
Mahalingam College of Engineering and Technology, segmentation models include Medial Axis, Reeb Graph,
Pollachi, India Mesh contraction and Volumetric contraction algorithms.
3
Department of Artificial Intelligence Engineering, Research
Center for AI and IoT, Near East University, Nicosia, Turkey

123
17784 Neural Computing and Applications (2022) 34:17783–17792

This paper proposed Intelligent Fuzzy based Clustering clustering for segmenting the faces into parts where a seed
scheme (HE-FCM and HE-IFCM) to cluster the vertices is chosen to represent the centers to determine the closest
and to embed the watermark in the 3D models based on faces to be assigned to the individual segments. The seed
surface based segmentation. Intelligent Fuzzy based Clus- value is recomputed to minimize the distance among the
tering scheme embeds the bits in every cluster using faces in every segment until it converges.
position identification strategy resulting in high capacity A robust watermarking scheme proposed by El Zein
embedding. This scheme enhances the robustness and et al. [12] embeds and extract watermark from vertices
imperceptibility and is compared with the existing algo- selected using Fuzzy C Means clustering without causing
rithm. The remainder of this paper is structured as follows. perceptual deformation. Aghila et al. [13] embedded up to
Section 2 describes the related works. Pre-processing, nine watermarking bits in each vertex of part of re-trian-
clustering, embedding and watermark extraction from 3D gulate mesh without causing much distortion in visual
models in the proposed work is explained in Sect. 3. Sec- quality and the geometric properties of the host model.
tion 4 illustrates the experimental analysis, results and Soliman et al. [14] selected vertex coordinates using
discussion followed by conclusions in Sect. 5. K-means clustering and converted to spherical coordinates
where the watermark information is embedded at random
bit positions. The optimal vertex location with minimum
2 Related work distortion is selected using minimum fitness function with
roulette wheel selector, linear crossover and Gaussian
Mesh segmentation can be performed by various algo- mutation as the basic operators. Mouhamed et al. [15]
rithms like convex decomposition [1], Fuzzy clustering [2], identified vertices to embed the watermark using K-means
K-Means [3], Hierarchical clustering [4, 5], Spectral clustering and derived the optimum group using particle
Clustering [6], Critical Point analysis [7], Random walks swarm optimization. From the clustered group, point of
[8], and Randomized cuts [9]. Segmenting 3D models interest is selected and watermark is embedded in the
based on convex shape depends upon the local and global spherical coordinates and achieved good imperceptibility
surface properties. The smaller surface properties are and robustness.
grouped together to form larger patches. Such schemes are Motwani [16] proposed a non-blind watermarking
termed as greedy or local clustering approaches. Region scheme using Support Vector Machine (SVM). Binary
growing is the best method for surface segmentation. A set SVM classifier is used to classify the vertices as appro-
of seeds are selected and for each seed the region are grown priate and inappropriate for watermark insertion with the
until all similar regions are grouped into a single region. help of genetic algorithm. Optimized clusters using seg-
The hierarchical regions try to merge the pair of regions mentation algorithms group’s similar vertices of 3D mod-
from bottom to top or from top to bottom in any hierarchy. els. Embedding the watermark into these similar vertices
For every iteration similar adjacent regions with minimal with appropriate factors might lead to low distortion. Same
error are merged and are repeated until the stopping cri- data is embedded repeatedly in different clusters resulting
terion is reached. in high capacity, overcoming cropping attack.
Attene et al. [10] described a hierarchical face clustering Both spatial and transformed domains [17–20] has
algorithm from triangular meshes by generating the binary several research works for 3D models in steganography and
tree of clusters. Every triangle is considered as a single watermarking schemes. Spatial domain [21–26] supports
cluster and after each iteration; all the pairs of adjacent large embedding capacity but lacks in robustness and poor
clusters are combined. These clusters are approximated and distortion whereas transformed domain supports better
the errors are evaluated to choose the most elite primitive robustness. Cayre et al. [24] proposed an algorithm based
to approximate the suitable set of clusters. on substitution procedure which can overcome similarity
Tan-Chi et al. [11] proposed a hierarchical segmentation and vertex reordering attacks. The main drawback of this
by decomposing the polygonal objects into meaningful scheme is its low embedding capacity and inefficient
parts based on the higher degree of salience. The degree of traversal of triangle. Bogomjakov et al. [18] embedded the
salience is same at each level of hierarchy and is higher watermark through reordering the vertices resulting in less
than the descendant parts. This scheme automatically embedding capacity. Later Wang et al. [25] improved the
determines the boundaries of each level in hierarchy and is above scheme for better embedding capacity through
based on the minimum slice perimeter that represents the advanced jump strategy via multi-level embedding pro-
non-local shape features for object interpretation. gress. A multi layered embedding scheme proposed by
K-means algorithm is is used for partitioning and seg- Chao [26] achieved high embedding capacity but the major
menting the 3D mesh models by clustering the coefficients. limitation of this scheme is as the embedding capacity
Golovinskiy [9] and Shlafman [3] performed K-means increases distortion level also increases.

123
Neural Computing and Applications (2022) 34:17783–17792 17785

Garg et al. [27] identified less visual distortion when space. An edge eij is a pair of connected vertices Vi and Vj.
embedding the watermark in vertices of deep surfaces than A face can be formed with three vertices (triangular mesh),
vertices of flat and peek regions. The watermark embed- four vertices (quads), or with n vertices (n-gons). In tri-
ding is done by re-positioning the selected vertices from angle mesh each face Fi =(Vi, Vj, Vk) consists of three
their original positions according to the category they vertices that form a triangle shape as shown in Eq. (1).
belong to. Yang et al. [21] embedded watermark in least 2 3 2 3 2 3
v1 f1 vi1 vj1 vk1
significant bits using quantization scheme and achieved 6v 7 6 f 7 6 vi2 vj2 vk2 7
high embedding and low distortion but couldn’t withstand 6 27 6 27 6 7
V ¼6 7
6 .. 7 F ¼ 6 . 7¼6
6 . 7 6 .. ..
7
.. 7 ð1Þ
any attacks. An adaptive embedding scheme proposed by 4 . 5 4 . 5 4 . . . 5
Tsai et al. [22] identified the complexity estimations of the
vm fm vin vjn vkn
spatial values. Cheng et al. [20] embedded six bits per
vertex using rearrangement. Lin et al. [19] embedded where 1  i, j, k  m.
watermark data in representation domain and achieved low
embedding capacity without distortion. Various research 3.2 Intelligent Fuzzy based clustering
results show that increase in embedding capacity results in scheme for 3D model
high distortion. So attaining high embedding capacity with
minimal distortion is the major challenge in 3D water- Initially polygonal mesh is preprocessed using principal
marking schemes. component analysis (PCA) to obtain a standard model for
embedding process. The clustering phase in the proposed
model is carried out with the preprocessed data. The
3 Proposed work standard Intelligent Fuzzy based clustering approach called
Fuzzy C Means is applied for clustering the vertices of the
Fuzzy clustering based embedding scheme preprocess the 3D model. Further the clustering is enhanced using Intel-
3D model using Principle Component Analysis (PCA) to ligent Fuzzy C Means algorithm and the results are com-
obtain a normalized 3D model. Vertices of the 3D model pared to find the best base model for identifying the
are clustered using Fuzzy C means clustering algorithm. interval to achieve high embedding capacity in 3D model.
HE-IFCM embeds the bits in every cluster using position
identification strategy resulting in high capacity embed- 3.2.1 Standard Fuzzy C- means method
ding. The embedded watermark can be extracted by per-
forming the reverse operation. An intelligent fuzzy The Fuzzy C-Means (FCM) algorithm was designed by
clustering scheme called Fuzzy C-means algorithm and an Bezdek et al. [31]. Initial process starts with a random
enhanced version called Intelligent Fuzzy C-means algo- guess to pick the centroids of every cluster. The algorithm
rithm are applied for clustering the vertices of the 3D tries to associate each data point into one cluster based on
model. Overview of the High Embedding Capacity in 3D the likelihood of the data point. This performs an iterative
Model using intelligent fuzzy clustering scheme is shown process that tries to minimize the objective function where
in Fig. 1. every data is assigned to the respective clusters. The higher
degree value is assigned to the data items that have short
3.1 3D Model representation distance from cluster centroids and lower to those that have
large distance from that of cluster centroids. After each
Geometric 3D models include spatial relationships, par- iteration, the cluster centroids are revised as per the
ticularly analytical geometry, where algebraic forms objective function given in Eq. (2).
expressed as relationships. Coordinates of the 3D models N X
X C  2
are termed as axes labelled x, y and z representing the JFCM ¼ Uijm Xi  Vj ð2Þ
length, breadth and height respectively. 3D models [28–30] i¼1 j¼1
can be represented as mesh or polygons where mesh is
where N ! Total items of the data, C ! Number of
described by vertices, edges and faces. In computer aided
clusters, Vj ! Centroid of cluster j, U ! Membership
design, 3D object are usually represented as triangle or
matrix which is computed as given in Eq. (3) and the
polygonal mesh.
cluster centroid is computed as given in Eq. (4).
A polygonal mesh is a triple T = {V, E, F} with m
vertices V= {v1 , v2 , ...vm}, set of edges E = {eij }, 1  i,j 
m and n faces F={f1 , f2 , ...fn }. Each vertex Vi = (xi,yi,zi )
consists of three values representing the position in 3D

123
17786 Neural Computing and Applications (2022) 34:17783–17792

Fig. 1 Overview of the High

capacity embedding scheme

1 center of the cluster are given a higher degree value. After

Uij ¼  m1
Pc Xi Vj
2
ð3Þ each iteration the cluster centers are revised.
k¼1 Xi Vk Consider there are n data points, V= {V1 ; V2 ; :::::Vn }.
PN Among this, n data points randomly choose K number of
i¼1 Uijm Xi
Vij ¼ PN ð4Þ data points as Cluster Centers (CC). CC=
i¼1 Uijm {CC1 ; CC2 ::::::CCK }.
During the initial process the membership degree m for Step 1 The membership matrix (W) is randomly
data point c against cluster j is initialized with a random initialized according to the Eq. (5) to satisfy the
number. implementation of fuzzy partitioning.
Xn
3.2.2 Intelligent Fuzzy C-means Clustering Wij ¼ 1; 8j ¼ 1; 2; :::; k ð5Þ
i¼1

Basically, Fuzzy C-Means (FCM) clustering algorithm is Step 2 The objective function of modified-FCM algo-
an enhancement of K-means algorithm, where the data rithm is given in Eq. (6)
points are associated with more than one cluster center with
Xm X K  
different membership values between 0 and 1. FCM’s ObjFCM ¼ Wijm Vi  New CCj2 ð6Þ
objective is to find cluster center that minimize the position i¼1 j¼1
of dissimilarity. The algorithm starts from the given data
points with an initial random selection of chosen cluster where m is the fuzzy coefficient.
centers. The FCM algorithm attempts to connect each of Step 3 Calculate new cluster centers using the following
the data points to one of the clusters by assigning a degree Eq. (7).
Pn n
of membership (W) based on probability. In clusters, a i¼1 Wij Vi
probability distribution function is used for assigning New CCj ¼ Pn m ð7Þ
i¼1 Wij
degree of membership for each data point. The Intelligent-
FCM is an iterative approach that seeks to minimize Step 4 Calculate similarity measure between the data
objective function. The objective function is minimized points and the cluster center using Euclidean distance as
and the algorithm terminates when each data point gets given in Eq. (8).
closer to their respective cluster centers. Those objects
dðVi ; CCj Þ ¼ jjVi  CCj jj ð8Þ
which have a short distance from the center of the cluster
and lower to those that have a large distance from the

123
Neural Computing and Applications (2022) 34:17783–17792 17787

Step 5 For the data point Vi , perform membership and the next 23 bits refers the mantissa part. This 32 bit
updation using Eq. (9), where l is the number of floating point representation allows us to represent values
iterations. between 1.1175494e-38 and 3.402823e?38 with decimal
1 precision of up to 6 to 7 places. The watermark data can be
Wij ¼
Pk jVi CCj jm1
2
ð9Þ embedded in the mantissa part of the vertex coordinates as
j¼1 jVi CCl j shown in the Fig. 3.
Mantissa part includes the bits from 10 to 31. watermark
As given in Eq. (10), the membership function is further bits are not embedded at the Least Significant Bit (LSB)
modified using a membership co-efficient q whose value regions of the mantissa part because during conversion of
is between 0.01 and 0.05. bits to floating values the embedded value is not reflecting
Wijtþ1 ¼ ð1  qÞWijt ð10Þ the modification. So bit embedding has to be performed at
the MSB region of the mantissa. ‘A’ refers to the number of
The modified membership function helps in the quick bits to be embedded and ‘R’ refers to the position of the
convergence of FCM than the original version. selected region. If A = 3, then 3 bits are embedded in the
n o
Step 6 Stop when maxij Wijlþ1  Wijl \C where C is a MSB region of the x, y, z coordinates of the vertex. If the
data is embedded in the bit range 25 to 31, retrieval of the
very low permissible value.
data is not possible during conversion decimal round off
The objective function is minimized in terms of optimal value. Data reconstruction is high for the watermark
cluster formation. In each iteration new cluster centers embedded between bits 10 to 14 but results with very high
have been identified to minimize the similarity measure visual distortion. Data embedded in the bit range 15–24,
between the data points and cluster center. The algorithm generated models with less visual distortion along with
terminates when each data point gets closer to their efficient reconstruction of watermark. The range of bits to
respective cluster centers. Those objects which have a short be embedded is identified by experimenting with good
distance from the center of the cluster and lower to those number of 3D triangular mesh models.
that have a large distance from the center of the cluster are
given a higher membership value. This improved mem-
bership value results in high embedding capacity in the 4 Results and discussions
host model.The main contribution of High Embedding
Intelligent Fuzzy C Means (HE-IFCM) algorithm is to
analyze FCM based watermark embedding by shifting the HE-IFCM implemented in python and analyzed for
values at certain interval in the host 3D model. This various 3D models. The 3D models like Bunny, Horse,
scheme guarantees that the embedding distortion is less and Venus are used for experiments are shown in the Fig. 4. 3D
does not get distracted much with the increase in the models are normalized and preprocessed using PCA. Since
embedded data. 3D models includes numerical, positive and negative
numbers, this normalization and pre-processing resulted in
3.3 Embedding process a standard range of values enhancing the embedding pro-
cess. The number of clusters to be formed using modified
The selection of vertex bit for embedding the watermark FCM identified using elbow model. From the plot obtained
and generation of watermarked 3D model is as follows. using elbow method, the variations are considered as the
The vertex set V includes real numbers, decimal numbers optimal number of clusters to be selected. The elbow curve
and signed numbers. The coordinates of the vertex is rep- obtained for the Bunny 3D model to find the optimum
resented using IEEE 754 single precision floating point number of clusters required for the clustering method is
representation. The 32 bit representation has three distinct given in the Fig. 5. The red circle indicates the optimum
part, sign, exponent and mantissa as shown in the Fig. 2. clusters identified as 3. The watermark data is embedded
The Most Significant Bit (MSB) represents sign bit into the selected bits of each cluster. The watermark data is
where positive sign is represented as 0 and negative sign as embedded into the MSB region of the mantissa and not in
1. The following 8 bit values represent the exponent part the LSB region.

Fig. 2 Representation of floating value by IEEE 754 Normalization Fig. 3 Selected bits for embedding the watermark data

123
17788 Neural Computing and Applications (2022) 34:17783–17792

Fig. 4 3D models used in the

experiments

Error (RMSE), Correlation and Peak Signal to Noise Ratio

(PSNR).

4.1.1 Hausdorff distance

Hausdorff distance estimates the extent to which each point

of an object set lies near some point of another object set
and vice versa [32] and is computed using Eq. (11). This
distance estimates the degree of resemblance between each
vertex v of host model M and watermarked model M 0 .
DðM; M 0 Þ ¼ maxðdH ðM; M 0 Þ; dH ðM 0 ; MÞÞ ð11Þ

Fig. 5 Clusters identified using elbow method Hausdorff distance computed by embedding different
number of bits at different positions for models like Bunny,
Vertices of the 3D model are clustered using modified Horse and Venus are shown in Figs. 8, 9, and 10. The
FCM algorithm. Based on the membership of each data results shows that dissimilarity obtained is negligible up to
point they are grouped into clusters with optimized 9 bits per vertices and results in higher dissimilarity from
objective function. All the vertices that are clustered into 10 to 15 bits per vertex.
has equal similarity value with respect to distance and The Hausdorff distance of the HEFCM is compared with
connectivity. Number of clusters k is selected using the various existing models and is shown in Table 1. Hausdorff
elbow plot. 3D models are preprocessed using PCA before distance attained for various models ranged from 5:089 
clustering and clusters generated by modified FCM are 105 to 6:0234  105 . The resemblance between the
shown in Fig. 6. watermarked and original host model is high for the
Results have been presented in Fig. 7 to show the visual HEMFCM. The resultant Hausdorff distance for the pro-
distortion of the 3D model after embedding 3 bits (one in posed model is lesser than that of [33, 13], and [27].
each of x, y, z coordinates of the 3D model) at different HEMFCM method attained a correlation coefficient 1
positions. Fuzzy C Means algorithm generated six clusters for most of the models same as like [33, 13], and [27]. The
for the Horse model whereas the proposed model resulted RMSE and PSNR of the HEMFCM is compared with
with three optimised clusters. Watermark is embedded into various existing models are shown in Tables 2 and 3.
the generated clusters. This optimised clusters resulted in
high embedding and low degradation effects in the water- 4.1.2 Peak Signal to Noise Ratio
marked model.
PSNR is an expression for the ratio between the maximum
4.1 Performance evaluation possible value (power) of a signal and the power of dis-
torting noise that affects the quality of its representation
Various 3D models are analyzed using HE-IFCM and is given in Eq. (12).
scheme to show the embedding capacity, quality of the  
Dmax
watermarked model in terms of human visualization and PSNR ¼ 20log ð12Þ
RMSE
fidelity measures. The watermarking techniques are eval-
uated based on Hausdorff distance, Root Mean Square where Dmax is the diagonal distance between the host mesh
model and the watermarked model.

123
Neural Computing and Applications (2022) 34:17783–17792 17789

Fig. 6 Optimized Fuzzy C

means clustering

Fig. 7 Results after embedding

data into different bit positions

123
17790 Neural Computing and Applications (2022) 34:17783–17792

Fig. 8 Hausdorff distance

computed by embedding bits 3,
6, 9, 12 in Bunny

Fig. 9 Hausdorff distance

computed by embedding bits 3,
6, 9, 12 in Horse

Fig. 10 Hausdorff distance

computed by embedding bits 3,
6, 9, 12 in Rabbit

4.1.3 Embedding capacity Table 1 Hausdorff distance for different schemes

Embedding capacity of the model can be measured using Model Sharma [33] Aghila [13] Garg [27] Proposed
the embedding rate (ER) Bunny 0:49  105 0:50  105 4:12  105 5:01  105
EC ¼ 3  kjVj ð13Þ Dragon 0:01  105 5:20  105 3:25  105 6:02  105
5 5 5
Horse 0:01  10 0:20  10 3:21  10 4:01  105
EC
ER ¼ ð14Þ Rabbit 0:01  105
5:80  105
4:71  10 5
5:09  105
jVj
15
where 0  k\ log210 1 þ1 is the quantity of the embedded
information in X, Y, and Z coordinates of the vertices. The shows that the PSNR obtained for various models is above
embedded data per vertex for various models are shown in 55dB even after embedding 15 bits per vertices.
Table 4.
The distortion rate computation using PSNR shows that
the HEMFCM model results with minimal distortion. The 5 Conclusion and future scope
PSNR value obtained for models like Bunny, Rabbit, Horse
is around 69.25 dB, 78.46 dB, and 73.12 dB respectively. HE-IFCM embeds the watermark into the 3D model by
The similarity between the host and watermarked model is partitioning the vertices with similar characteristics into
high after embedding is proved using obtained PNSR and clusters using modified Fuzzy C Means algorithm. The
RMSE. The embedding capacity is measured for bits per watermark data is embedded into the selected bit positions
vertices with respect to the number of vertices. The results of the vertices in spatial domain. The embedding capacity,
and distortion rate is analyzed using PSNR, and Hausdorff

123
Neural Computing and Applications (2022) 34:17783–17792 17791

Table 2 Root Mean Square

Model Sharma [33] Aghila [13] Garg [27] Zein [12] Proposed
Error for different schemes
Bunny 0:14  105 0:19  105 0:32  104 0:31  103 0:17  106
2 3
Horse 0:1  10 5
0:12  10 5 0:26  10 0:31  10 0:12  106
Rabbit 0:91  105 0:12  105 0:36  104 0:11  103 0:23  105
4 3
Dragon 0:82  10 5
0:59  10 5 0:49  10 0:20  10 0:21  106

tortion. HEMFCM is successful in embedding large

Table 3 Peak Signal to Noise Ratio for different schemes capacity of watermark resulting in minimal distortion.
Model Aghila [13] Gard [27] Proposed The detection of geometric salient points using opti-
mization can be analyzed for high data embedding and
Bunny 55.34 66.75 69.25
retrieval. Since 3D model includes faces with three or four
Dragon 48.56 52.76 64.21
edges, plane partitioning can be performed to cluster the
Rabbit 55.70 75.32 78.46
vertices for better segmentation. Visible watermarking in
Horse 56.59 68.34 73.12
3D model has not attained much research attention due to
the large deviation shown during embedding. Hence the
extension of our proposed watermarking can be analysed
with visible watermarking technique.
distance. This scheme is capable of embedding 3 KB to 15
KB for small models and 164 KB to 820 KB for larger
models achieving around 68 dB of PSNR with less dis-

Table 4 Embedding capacity

Model # Vertices Bits per vertices Embedded data(bits) PSNR
and PSNR at different
embedding rates Horse 8431 3 25293 67.68
6 50586 65.58
9 75879 64.52
12 101172 58.42
15 126465 56.21
Venus 134345 3 403035 73.76
6 806070 72.21
9 1209105 68.39
12 1612140 65.27
15 2015175 63.11
Dragon 437645 3 1312935 67.78
6 2625870 6 6.65
9 3938805 64.21
12 5251740 62.34
15 6564675 59.11

123
17792 Neural Computing and Applications (2022) 34:17783–17792

Declarations 16. Motwani RC, Motwani MC, Bryant BD, Harris Jr FC, Agarwal
AS (2010) ‘‘Watermark embedder optimization for 3d mesh
Conflict of interest The authors Modigari Narendra, M L Valarmathi, objects using classification based approach, In: 2010 International
L Jani Anbarasi, Vergin Raja Sarobin M, and Fadi Al-Turjman Conference on Signal Acquisition and Processing, pp 125–129,
declare that there is no conflict of interest. IEEE
17. Petitcolas FA, Anderson RJ, Kuhn MG (1999) Information hid-
ing-a survey. Proc IEEE 87(7):1062–1078
18. Bogomjakov A, Gotsman C, Isenburg M (2008) Distortion-free
References steganography for polygonal meshes. Comput Gr Forum
27:637–642
1. Chazelle B, Dobkin DP, Shouraboura N, Tal A (1997) Strategies 19. Lin C-H, Chao M-W, Chen J-Y, Yu C-W, Hsu W-Y (2013) A
for polyhedral surface decomposition: an experimental study. high-capacity distortion-free information hiding algorithm for 3D
Comput Geometry 7(5–6):327–342 polygon models. Int J Innov Comput Inf Control 9(3):1321–1335
2. Katz S, Tal A (2003) Hierarchical mesh decomposition using 20. Cheng Y-M, Wang C-M (2006) A high-capacity steganographic
fuzzy clustering and cuts. ACM Trans Gr (TOG) 22(3):954–961 approach for 3D polygonal meshes. Vis Comput 22(9):845–855
3. Shlafman S, Tal A, Katz S (2002) Metamorphosis of polyhedral 21. Yang Y, Peyerimhoff N, Ivrissimtzis I (2012) Linear correlations
surfaces using decomposition. Comput Gr Forum 21:219–228 between spatial and normal noise in triangle meshes. IEEE Trans
4. Garland M, Willmott A, Heckbert PS (2001) ‘‘Hierarchical face Visual Comput Gr 19(1):45–55
clustering on polygonal surfaces,’’ In: Proceedings of the 2001 22. Tsai Y-Y (2014) An adaptive steganographic algorithm for 3D
symposium on Interactive 3D graphics, pp 49–58 polygonal models using vertex decimation. Multimed Tools Appl
5. Gelfand N, Guibas LJ (2004) ‘‘Shape segmentation using local 69(3):859–876
slippage analysis,’’ In Proceedings of the 2004 Eurographics/ 23. Ohbuchi R, Masuda H, Aono M (1998) Watermarking three-
ACM SIGGRAPH symposium on Geometry processing, dimensional polygonal models through geometric and topological
pp 214–223 modifications. IEEE J Sel Areas Commun 16(4):551–560
6. Liu R, Zhang H (2004) ‘‘Segmentation of 3d meshes through 24. Cayre F, Macq B (2003) Data hiding on 3-D triangle meshes.
spectral clustering,’’ In: 12th Pacific Conference on Computer IEEE Trans Signal Process 51(4):939–949
Graphics and Applications, 2004. PG 2004. Proceedings. 25. Wang C-M, Cheng Y-M (2005) An efficient information hiding
pp 298–305, IEEE algorithm for polygon models. Comput Gr Forum 24:591–600
7. Lin H-YS, Liao H-YM, Lin J-C (2006) Visual salience-guided 26. Chao M-W, Lin C-H, Yu C-W, Lee T-Y (2008) A high capacity
mesh decomposition. IEEE Trans Multimedia 9(1):46–57 3D steganography algorithm. IEEE Trans Visual Comput
8. Lai Y-K, Hu S-M, Martin RR, Rosin PL (2008) ‘‘Fast mesh Graphics 15(2):274–284
segmentation using random walks,’’ In: Proceedings of the 2008 27. Garg H, Agarwal S (2013) A secure image based watermarking
ACM symposium on Solid and physical modeling, pp 183–191 for 3D polygon mesh. Sci Technol 16(4):287–303
9. Golovinskiy A, Funkhouser T (2008) ‘‘Randomized cuts for 3d 28. Jian-qiu J, Min-ya D, Hu-jun B, Qun-sheng P (2004) Water-
mesh analysis,’’ In: ACM SIGGRAPH Asia 2008 papers, marking on 3D mesh based on spherical wavelet transform.
pp 1–12 J Zhejiang Univ-Sci A 5(3):251–258
10. Attene M, Falcidieno B, Spagnuolo M (2006) Hierarchical mesh 29. Kwon K-R, Kwon S-G, Lee S-H, Kim T-S, Lee K-I (2003)
segmentation based on fitting primitives. Vis Comput Watermarking for 3d polygonal meshes using normal vector
22(3):181–193 distributions of each patch, In: Proceedings 2003 International
11. Ho T-C, Chuang J-H et al (2012) Volume based mesh segmen- Conference on Image Processing (Cat. No. 03CH37429), vol 2,
tation. J Inform Sci Eng 28(4):705–722 pp II–499, IEEE
12. El Zein OM, El Bakrawy LM, Ghali NI (2017) A robust 3d mesh 30. Karni Z, Gotsman C (2000) Spectral compression of mesh
watermarking algorithm utilizing fuzzy c-means clustering. geometry, In: Proceedings of the 27th annual conference on
Future Comput Informat J 2(2):148–156 Computer graphics and interactive techniques, pp 279–286
13. Thiyagarajan P, Natarajan V, Aghila G, Venkatesan VP, Anitha R 31. Bezdek JC, Ehrlich R, Full W (1984) Fcm: the fuzzy c-means
(2013) Pattern based image steganography. 3D Res 4(1):1 clustering algorithm. Comput Geosci 10(2–3):191–203
14. Soliman MM, Hassanien AE, Onsi HM (2015) A Robust 3D 32. Atallah MJ (1983) A linear time algorithm for the hausdorff
Mesh Watermarking Approach Using Genetic Algorithms. In: distance between convex polygons, Department of computer
Intelligent Systems’2014. Advances in Intelligent Systems and science technical reports
Computing, vol 323. Springer, Cham, pp 731–741. https://doi. 33. Sharma LK, Ojha DB (2015) Imperceptible watermarking
org/10.1007/978-3-319-11310-4_64 scheme for 3D triangular mesh. Int J Appl Innov Eng Manag
15. Mouhamed MR, Soliman MM, Darwish AA, Hassanien AE 4(12):28–34
(2018) Interest points detection of 3d mesh model using k means
and shape curvature, In: International Conference on Advanced Publisher’s Note Springer Nature remains neutral with regard to
Intelligent Systems and Informatics, pp 415–424, Springer jurisdictional claims in published maps and institutional affiliations.

123