Machine Learning Based Footprint Recognition
Machine Learning Based Footprint Recognition
Machine Learning Based Footprint Recognition
ANSHU GUPTA
Department of Computer Science, Babasaheb Bhimrao Ambedkar University, Lucknow (U.P.), India.
Email: [email protected]
DEEPA RAJ
Department of Computer Science, Babasaheb Bhimrao Ambedkar University, Lucknow (U.P.), India.
Email: [email protected]
Abstract
With the escalating digitization of the modern era and the infusion of technology into almost every aspect
of human lives, biometric authentication systems have become the need of the hour to control the identity
thefts. They are the automated systems that divulge individuals identify based on their unique biological or
behavioral personality traits like fingerprints, face, iris etc. This article presents personal recognition using
Foot biometrics by following machine learning approach. The implementation of three supervised machine
learning methods namely, Regression, Classification and ANN (artificial neural network) has been done.
The proposed method works in two stages: Geometric Feature Extraction and implementation of Machine
Learning algorithms. Firstly, handcrafted foot features are extracted using geometrical methods which are
instilled as input to three supervised machine learning algorithms to predict the identity of user.
Experimental results reveal that the weighted KNN model is the most performant method among all the
implemented classifiers with 99.5% validation accuracy and the overall training time of 0.29886 seconds.
While other two utilized and tested supervised machine learning methods, also achieved a reasonable
accuracy of 99.15% by Squared Exponential GPR model (Regression) and 97.47% by ANN (Feed Forward
Neural Network with Back Propagation).
Keywords: Biometrics, Footprint Recognition, Feature Extraction, Geometrical Features, Machine
learning, ANN, Regression, Classification
INTRODUCTION
In this technology driven era, the profusion of network security breaches and identity
thefts, make a tremendous stipulation for sturdy and robust biometric authentication
systems that truly rely on something that “you are” with unique God gifted physiological
or behavioral characteristics as opposed to something “you know”(passwords, IDs etc.)
or something “you have”(Identity cards, keys etc.). To have an easy and secure access
to the digital world, many biometric systems like face, finger, iris, gait and ear, are
prevalent. Unlike other biometrics, Foot Biometry has achieved limited global acceptance.
This very idea is the catalyzation and motivation behind present research work to explore
new potentials in foot biometrics comparable to state-of-the-art biometric techniques.
Footprint identification can be applied in many areas like protection against child thefts,
forensics, defense systems and legal capacitance etc. This paper proposes a novel
archetype scheme for footprint recognition by amalgamating feature extraction and
machine learning algorithms. The working proposal can be broadly divided into two parts;
drilling down Geometrical foot Features followed by the application of supervised machine
learning techniques. At first, 20 hand crafted geometric foot features are extracted from
each sample footprint image which are fed into the machine learning algorithms as an
input foot feature vectors in the second part of proposed technique. In this paper, three
supervised machine learning methods namely regression, classification and ANNs have
been applied for predicting the recognition outcomes. To measure the effectiveness of
proposed algorithms, the biometric performance is tested by calculating False Accept
Rate (FAR), False Reject Rate (FRR), Accuracy, Precision, Sensitivity, Specificity, F1-
Score and training time. The experimental results of recommended approach
demonstrate that the weighted KNN model (Classification) has achieved the highest
recognition accuracy of 99.5% (with 0.29886 seconds as an average training time)
compared to other two tested machine learning methods viz. Regression (99.15%) and
ANN (97.47%).
The rest of the paper is structured as follows: Section 2 briefly presents the related work
on footprint recognition. Section 3 presents the dataset used for experimentation. Section
4 introduces the proposed methodology of current work. Section 5 presents and analyzes
the experimental results. At the end, Section 6 concludes the paper with the hints to future
work.
2. RELATED WORK
This section provides a comprehensive literature available in the foot biometry
Robert B. Kennedy [1] can be awarded as the father of foot biometrics for exploring
human footprints for the first time as biometric feature and used them for medical and
forensic research by using inked bare footprints impressions. 38 local geometric features
were extracted by proposing 6 different methods and achieving the minimum error
recognition rates of 1.35% FMR and 2.18% FNMR. Nakajima et al. [2] performed a
conscientious study of pressure distribution of footprints using positional and directional
normalization and achieved recognition rate of 85% using Euclidean distance image
matching. A rotation invariant footprint based authentication system was introduced by
Uhl and Wild [3] who used geometry, shape and texture based foot features and
achieved with 97% recognition accuracy for 16 subjects.
To protect against the child pilfering, Jia et al. [4] proposed an automated footprint based
newborn personal recognition system by extracting 4 orientation based features (Ordinal
Code, Binary Orientation Co-occurrence Vector, Competitive Code and Robust Line
Orientation Code) and achieved 98% recognition accuracy. Also, Jia et al. [5] presented
another scheme based on band-limited phase-only correlation (BLPOC) for extracting
foot features using 101 footprint images of infants and claimed 97% recognition accuracy.
Infant’s footprints were further explored by Kotzerke et al. [6] who developed a novel
algorithm for feature extraction from crease patterns of infants at three diverse
timestamps i.e. three days since birth, eight weeks and six months. With the
implementation of morphological processing, directional filtering and block-wise crease
line reconstruction, they observed an EER of 22.22%.
Nagvanshi [7] and Kumar [8] incorporated various prevalent techniques like PCA, SOM,
SVD, HMM, NN, ART2 and Fuzzy logic etc. for footprint recognition. An extensive study
HOG, GLCM and LBP) in conjunction with the classification methods (namely, LDA, SVM,
KNN and ESD) using footprints and observed an accuracy of 97.9% for LBP + LDA as
the most performant scheme for personal verification.The empirical studies on footprint
biometrics done by Khokher and Singh [31] explored many footprint recognition
techniques along with their impediments and future dimensions. They indicated footprints
as a revealer of individuality information like age, height, gender, weight, health status
and region specific personality traits etc. A multi-model rank-level fusion system was
proposed by Kumar and Shekhar [32] by utilizing features from two modalities - palm
and foot with the suggested accuracy of 92-99% using fusion. Pataky [33] presented
correlation based approach by using 1040 dynamic plantar foot pressure images from
104 volunteers(40 males and 64 females)and obtained the classification rate of 99.6%.
Gupta and Raj [34-36] proposed a novel method for footprint recognition using eigenfeet
and a new distance metric for touch-less footprints with recognition accuracy of
97%.They further, presented a comparative analysis of texture based methods using
LBP, LPQ, SIFT and SURF methods using footprints and claimed a 98% recognition
accuracy. They also introduced an innovative approach for footprint biometrics using
chronological implementation of LBP and SIFT with the recognition accuracy of 98.93%.
3. DATASET
Figure 1. Sample footprint images of 10 subjects from Nagwanshi and Dubey [37]
database
Presently, there are only two publically available databases of human footprint images for
biometric experimentation. The first database is uploaded by Nagwanshi and Dubey
[37] in the IEEE Dataport open access repository of scanned grayscale planter footprint
images of left feet of 220 volunteers. By varying the hue and saturation levels at different
times, 6 diverse image samples are captured for each person, hence, add up to a total of
1320 (=6X220) images. The size of each image is 256 X 666.Another dataset has been
uploaded by R. Kumar [38] in 2019 at GitHub repository, now available at kaggle. It is a
database of 100 scanned plantar footprint images of 21 persons and 100 dactyloscopic
images (including left and right footprints) of 32 individuals with two to five images per
subject [37].The first database [37] of 1320 images is being used in this paper for
experimentation. Following Figure 1 shows the sample footprint images of ten persons.
4. PROPOSED METHODOLOGY
In this section, the detailed description of proposed prototype scheme of foot biometrics
is exhibited. Following Figure 2 shows the basic schematic diagram of proposed
framework. Working methodology has been broadly divided into two phases:
1. Generation of Handcrafted Geometric foot features using feature extraction
2. Generation of Machine Learning model for foot recognition
Figure 2. Proposed Framework
I. Phase 1
Phase 1 of the working methodology starts with the pre-processing steps followed by
outer foot helix detection of registered plantar footprint images. Afterwards, geometric
feature extraction followed by dimensionality reduction and normalization of feature
vectors during enrollment and authentication process. The resulting dataset of extracted
geometric feature vectors become the input vectors for the implementation of machine
learning model in phase 2. Diagrammatically, various steps of phase 1 are shown below:
Figure 3. Phase 1 of proposed method
The detailed description of each step followed in phase 1 of the proposed scheme is
explained as below. Table 1 illustrates the listing of main abbreviations used in the
following steps.
a) Pre-processing and the foot helix detection
The presence of noise in the scanned footprint images can hinder the results of obtaining
outer helix of the footprint images. Hence, morphological operation-dilation is performed
by using disc as structuring element resulting in the enlarged foot boundaries and small
holes in objects are filled. After the application of dilation, outer helix of foot is detected
using canny edge operator. Figure 4 shows the typical edge detection results to extract
the foot contours with/without dilation. It can easily be observed that by applying dilation
followed by edge detection, improved results are obtained as cluttering edges and noise
are suppressed and salient contours are enhanced.
enough to characterize the contour of outer helix of foot for personal identification and
hence, achieve a better recognition performance.
Algorithm 1
Input: Binary edge (boundary) of footprint image.
Output: Ut and Lh , Lw and Rw, FHLmax and FWLmax as feature vectors FV1 and FV2
1. Search among the boundary points having minimum y value to get the uppermost toe
point Ut.
2. Similarly, find Lh, the lowermost heel point, as the boundary point having maximum y
value.
3. Also, find the boundary points Lw and Rwwith the minimum and maximum x values
respectively as the leftmost and rightmost endpoints of FWL.
4. Calculate FHLmax by finding the length of the line joining endpoints Ut and Lh.
5. In the same way, find FWLmax by computing the length of the line joining Lw and R
Algorithm 3
Input: Ut and Lh ; Lw and Rw
Output: UPPER_ Area, LOWER_ Area, LEFT_ Area and RIGHT_ Area
1 Consider four triangles namely: UPPER_TRI, LOWER_TRI, LEFT_TRI and RIGHT_TRI
where
∆ UPPER_TRI→ Triangle formed with Ut , Lw and Rw
∆ LOWER_TRI→ Triangle formed with Lh, Lw and Rw
∆ LEFT_TRI→ Triangle formed with Lw ,Ut and Rw
∆ RIGHT_TRI→ Triangle formed with Rw, Ut and Lh
2 Find areas of step 1 four triangles as UPPER_ Area, LOWER_ Area, LEFT_ Area and
RIGHT_ Area respectively.
3 Designate UPPER_ Area, LOWER_ Area, LEFT_ Area and RIGHT_ Area as feature
vectors FV17, FV18, FV19 and FV20.
By following the above steps of algorithms 1, 2 and 3, a feature set FV= {FV 1, FV2... FV19,
FV20}, consisting of 20 geometric feature vectors, is extracted which characterizes the
shape of outer helix of foot for personal recognition. Finally, feature vector set consisting
of 26,400 features (=1320 ×20) from1320 plantar foot images gets generated. Table 2
illustrates the listing of sample feature vectors of 12 footprint images of 2 subjects.
in the range [0, 1] to achieve a normalized feature set. Table 3 lists 5 dominant features
of the corresponding 20 features of 12 footprint images of 2 subjects in Table 2.
Table 3. Corresponding reduced feature vectors by applying PCA
S.N. FV1 FV2 FV3 FV4 FV5
1 0.8105 0.1951 0.6266 0.6710 0.1968
2 0.6456 0.2729 0.6210 0.4752 0.1050
3 0.8026 0.1853 0.6287 0.6645 0.1860
4 0.7565 0.4517 0.6952 0.6139 0.3404
5 0.8649 0.4575 0.6405 0.7468 0.1179
6 0.7926 0.4546 0.6721 0.6410 0.2247
7 0.7189 0.4745 0.6527 0.6547 0.0806
8 0.7074 0.4629 0.6552 0.6535 0.0836
9 0.7183 0.4658 0.6539 0.6484 0.0711
10 0.7189 0.4745 0.6527 0.6547 0.0806
11 0.7130 0.4659 0.6543 0.6521 0.0790
12 0.7184 0.4680 0.6535 0.6503 0.0743
II. Phase 2
The normalized feature set of 6,600 features becomes the input for the application of
machine learning algorithms in phase 2 of proposed methodology. On the basis of
predicted output of machine learning algorithms, system decides that the claimed identity
is a genuine user or an imposter. Figure 7 shows diagrammatic representation of phase
2.
Figure 7. Phase 2 of proposed method
Basically, machine learning is the science of programming computers to learn from data
to automatically predict the results for new set of data. It uses two types of techniques:
supervised learning, which trains a model on known input and output data to generate
predictions in response to new data and unsupervised learning, which finds hidden
patterns or intrinsic structures in input data. Supervised learning uses classification and
regression techniques to develop predictive models.
Classification techniques predict categorical responses means Classification
models classify input data into categories. Typical applications include medical
imaging, image and speech recognition, and biometric identification etc.
EU_dx,y=√∑N
i=1(xi − yi )
2 (1)
Where EU_dx,y refers to euclidean distance, xi denotes input data, yi represents training
data, and N declares total number of features on the input data. Steps of K-NN can be
summarized as below:
1. Select the value of K as the number of the neighbors
2. Calculate the Euclidean distances of K neighbors with respect to query point
3. Select the K nearest / closest neighbors as per the calculated Euclidean distance.
4. Among K neighbors, the number of the data points in each class is counted i.e.
voting of neighbors
5. Assign the new data points to the class with the maximum number of the
neighbors.
b) Gaussian Process Regression (GPR) classifier
Gaussian process regression (GPR) model is kernel-based, nonparametric probabilistic
models. A GPR model is represented as:
y = h (x)T β + f(x), (2)
where,
h(x) = Set of basis functions transforming the original feature vector x in Rd into a new
feature
vector h(x) in Rp.
β = p × 1 vector of basis function coefficients.
f(x) = Gaussian Process with zero mean and covariance function, k (x, x′) i.e. GP (0,
k (x, x′) )
y = Observed response
The model of response y can be represented as:
P(yi ⎸f (xi), xi) ~ N (yi ⎸h(xi)Tβ + f(xi), σ2) (3)
In supervised learning method, the points with similar predictor values xi are expected to
have close response values yi. In Gaussian processes, the similarity is expressed using
covariance function which specifies the covariance between the two latent variables f
(xi) and f (xj), where both xi and xj are d × 1 vectors. The covariance function k(xi, xj) is
defined using various kernel functions based on kernel parameter vector θ. Hence,
covariance function can also be expressed as k(xi, xj ∣ θ). In general, the kernel
parameters depend on two factors; signal standard deviation σf and the characteristic
length scale σl which defines the minimum length between input values xi and response
values to be correlated.There are many built-in kernel (covariance) functions with same
length scale for each predictor are: Exponential Kernel, Squared Exponential Kernel,
Matern 3/2, Matern 5/2 and Rational Quadratic Kernel. Among them, squared
exponential covariance function is most commonly used kernel and is defined as:
1 (xi −xj )T (xi −xj )
k(xi , xj , ⎸Ɵ ) = σ2f exp [− ] (4)
2 σ2l
clustering raw input. By varying layers and neurons, NNs are modeled for classification
and recognition and many other applications [41]. In NNs , three types of layers are used:
input layer, hidden layer, and output layer. Every network has one input and one output
layer. All layers in between are referred to as hidden layers. The function that transforms
the input into the output is known as activation function e.g. Sigmoid, RELU, Tanh, Step,
Gaussian, Ramp and Linear etc.[42].The simplest mathematical representation for the
NNs using Sigmoid Activation Function can be summarized as:
1
ŷ = 1+e−z , where 𝑧=𝑥.𝑤+𝑏 (5)
Where 𝑥 represents the inputs vector, w refers to the weights vector, b is the bias and ŷ
denotes the predicted output.
Broadly, ANN is trained by using three approaches: Supervised Learning (Error based),
Unsupervised Learning and Reinforcement Learning (output based). In supervised
learning, there exists a mapping between input and target known output dataset and ANN
model is trained with the labeled dataset to predict the output while in Unsupervised
Learning, the target output is unknown, hence, ANN model learns on its own by
discovering hidden patterns in the input data [43]. Reinforcement learning is based on
rewarding desired behaviors and/or punishing undesired ones it enables to learn in an
interactive environment by trial and error using feedback from its own actions and
experiences. Our methodology uses supervised learning of ANN.
When ANN is trained using supervised learning, a known set of input data is presented
to the network model along with the known set of desired output data, which produces a
predicted output data. If there is a mismatch between predicted and desired output
values, an error signal is spawned which causes the weights adjustments until zero error
or minimum error value using back propagation Gradient Descent (GD) optimization
method. GD method calculates the Error Gradient with respect to all weights in the
network and is supplied to the optimization method to update the weights to minimize the
error. So, Differentiable activation functions are required to be used by Back propagation.
Mathematically, if ∆W ij is the weight update of link connecting the ith and jth neuron
of two neighboring layers, then ∆W ij is defined as:
∂E
∆W ij = η (− ∂W ) (6)
ij
∂E ∂E ̂
∂y ∂z
= ∂ŷ × ∂z × ∂W (8)
∂Wij ij
∂E ∂ 1
̂
= ∂ŷ (2 [y − ŷ]2 ) = −(y − ŷ)
∂y
̂
∂y ∂ 1 1 1
= ∂z (1+e−z ) = (1+e−z ) (1 − 1+e−z ) = ŷ(1 − ŷ)
∂z
∂z ∂
= ∂W (x. Wij + b) = x
∂Wij ij
∂E
⇒ = −(y − ŷ). ŷ(1 − ŷ). x
∂Wij
⇒Wnew=W old+ η(y − ŷ). ŷ(1 − ŷ). x (9)
So, the new weights are assigned / adjusted using equation 9 by propagating the error
backward through network and the process continues until error is minimized to an
acceptable low value or actual output is matched with the desired results.
5. RESULTS AND DISCUSSION
All the experiments are configured with Intel® Core™ i5-3110M CPU @ 2.40 GHz with 8
GB RAM, and NVIDIA® GeForce RTX™ 30 graphics card. Microsoft Windows 10 Pro 64-
bit operating system and MATLAB R2019a (64-bit) has been used for the experimentation
and evaluation of the proposed methodology.
5.1 Evaluation Metrics
To measure the effectiveness of proposed algorithm, the most widely adopted biometric
performance indicators are considered such as False Accept Rate (FAR), False Reject
Rate (FRR), Accuracy, Precision, Recall, Specificity and F1-score. These performance
metrics are computed using four primitive variables of confusion matrix i.e.TP, FP, FN
and TN. They are defined as below:
TP (True Positives) refers to the count of cases when system predicted positive
and it’s actually true.
FP (False Positives) means count of values when system predicted positive and
it’s actually false.
FN (False Negatives) gives the number of cases when system falsely predicted
the actual true values.
TN (True Negatives) counts the cases when system predicted negative and it’s
actually false.
It can be seen that actual results and predicted results contradict for FP and FN values.
So, errors are calculated using these values. Other performance matrices are defined as:
False Positive Rate (FPR or FAR): It is the rate at which system accepts
unauthorized users.
False Negative Rate (FNR or FRR): It is the rate at which system rejects
authorized/genuine users
Accuracy: refers to the fraction of correct predictions made by the ML Model.
Error Rate: It refers to the fraction of false predictions and equals to (1- Accuracy)
Precision: It measures quality of positive predictions made by the model means
fraction of relevant instances among positive retrieved instances
Recall (Sensitivity or TPR): It quantifies the model’s ability to detect positive
results means fraction of relevant instances among retrieved instances (positive or
negative).
F1-score: It is defined as the harmonic mean of model’s precision and recall.
From the above table, it is evident that weighted KNN model is outperforming among all
the mentioned classification models with 99.5% accuracy and overall training time of
0.29886 seconds. The confusion matrix, scatter plot and the corresponding ROC curve
of weighted KNN model are depicted in figure 8. It can be observed through the
experimental results that all the applied classification models are performing well. But
taking into consideration of training time, Decision Tree classifiers are the slowest ones
in spite of having reasonable accuracy. Ensemble methods are also following the same
scenario.
Figure 8a. Scatter Plot of Figure 8b. Confusion Matrix of Figure 8c. ROC curve of
Weighted KNN Weighted KNN Weighted KNN
B. Experiment 2
A comparison analysis is done by applying different regression models with an
implementation of 5-fold Cross-Validation strategy. Table 5 outlines the results of this
comparison study in terms of various performance metrics namely, RMSE (Root Mean
Squared Error), MSE (Mean Squared Error), MAE (Mean Absolute Error) and Training
Time in seconds.
Table 5. Comparison of various performance metrics of different classifiers for
Regression models
C. Experiment 3
It can be clearly seen that Squared Exponential GPR model is the best performing
model among all implemented regression models with the least RMSE value of 0.064363
and overall training time of 41.329 seconds. Figure 9 shows the corresponding graphs
of visualized results of Squared Exponential GPR model are shown below:
Figure 9c. Predicted vs. Actual
Figure 9a. Response Plot of Figure 9b. Confusion Matrix of
Plot of Squared Exponential
Squared Exponential GPR Squared Exponential GPR
GPR
This experiment seeks the efficacy of neural networks for footprint recognition. For the
purpose, Feed-forward back propagation Neural network is implemented with 10 neurons
at the hidden layer 1 with the Tansig as activation function and Purelin activation function
at hidden layer 2 (Figure 10a). Other parameters used are: Data Division – Random;
Adapting Learning Function- learnGDP; Network Training Function- Levenberg-
Marquardt to update weight and bias values; Performance Function- MSE The
experimental results show that the overall accuracy of 97.47% and the training time of
3.4358 seconds were achieved by applying ANN supervised learning approach. The
resulting statistics of other performance metrics are as follows:
Error Rate: 0.0252
Precision: 0.2857
Recall: 0.2500
Specificity: 0.9883
F1-scores: 0.2667
FPR: 0.0117
FNR: 0.7500
Following Figure 10b, c show the confusion matrix and performance plot of ANN:
Figure 10a. ANN Network Figure 10b. Confusion Matrix Figure 10c. Performance Plot
View of ANN of ANN
D. Experiment 4
From the results of previously discussed 3 supervised machine learning algorithms of
classification, regression and ANN, it was concluded that the weighted-KNN model is
the most performant method among all the mentioned methods with 99.5% validation
accuracy and the overall training time of 0.29886 seconds. Following Table 6 presents
the comparison chart of performance metrics of various machine learning algorithms used
in the proposed work. Also, Figure 11a and Figure 11b depict the bar charts of
Performance metrics and Training Times of the best performing above mentioned ML
algorithms i.e. Weighted-KNN, Squared Exponential GPR and ANN.
Table 6. Comparison of performance metrics of different classifiers
E. Experiment 5
Through the exhaustive literature study of foot biometry, it was revealed that only one
work from Basheer et al. [14] is available for footprint recognition that applies Machine
learning methods using fuzzy ensemble learning. They claimed the highest recognition
accuracy of 98.88%, FPR at 0.01, FNR at 0.093 and the training time of 18.510 seconds.
While, through the results of experiment 5 the proposed prototype scheme is
outperforming the [14] approach with 99.5% recognition accuracy, FPR at 0.00155039
and FNR at 0.27777778 with overall training time of 0.299 seconds, hence providing a
better robust solution for footprint recognition (Table 7 and Figure 12b).
Table 7. Comparison of performance metrics of proposed method with the
existing method [14]
Method Accuracy FPR FNR Training Time
time of 0.29886 seconds in comparison to Regression and ANN with the greatest
accuracy of 99.15% and 97.5% respectively. The application of optimized deep learning
techniques for footprint recognition is the next target of our present work extension. Also,
in near future, proposed scheme can be combined with other feature modalities to
develop a multimodal biometric system.
References
1. Kennedy RB. Uniqueness of bare feet and its use as a possible means of identification. Forensic
Science International, 82(1):81-87. (1996) DOI: 10.1016/0379-0738(96)01969-x.
2. Nakajima, K., Mizukami, Y., Tanaka, K. and Tamura, T., Footprint-based personal recognition. IEEE
Transactions on Biomedical Engineering, 47(11), pp.1534-1537(2000).doi: 10.1109/10.880106.
3. Uhl, A. and Wild, P. Footprint-based biometric verification. Journal of Electronic Imaging, 17(1),
p.011016(2008)
4. Jia, W., Cai, H.Y., Gui, J., Hu, R.X., Lei, Y.K. and Wang, X.F. Newborn footprint recognition using
orientation feature. Neural Computing and Applications, 21(8), pp.1855-1863.(2012)
doi:10.1007/s00521-011-0530-9
5. Jia, Wei & Hu, Rong-Xiang &Gui, Jie& Lei, Ying-Ke. (2010). Newborn Footprint Recognition Using
Band-Limited Phase-Only Correlation. 6165. 83-93. 10.1007/978-3-642-13923-9_9.
6. Kotzerke, J., Davis, S., Horadam, K., &McVernon, J. (2013, September). Newborn and infant footprint
crease pattern extraction. In 2013 IEEE International Conference on Image Processing (pp. 4181-
4185). IEEE.
7. Nagwanshi KK, Dubey S. Biometric authentication using human footprint. International Journal of
Applied Information Systems (IJAIS). 3(7):1-6. (2012)
8. Kumar, V.D.A., Ramakrishnan, M. Legacy of Footprints Recognition- A Review. International Journal
of Computer Applications, 35, 9–16.( 2011)
9. Khokher, Rohit & Singh, R C. (2016). Footprint-Based Personal Recognition using Scanning
Technique. Indian Journal of Science and Technology. 9. 10.17485/ijst/2016/v9i44/105167.
10. Hashem, K. M., & Ghali, F. (2016). Human identification using foot features. Int J EngManuf, 6(4), 22-
31.
11. Keatsamarn, T., & Pintavirooj, C. (2018, November). Footprint identification using deep learning. In
2018 11th Biomedical Engineering International Conference (BMEiCON) (pp. 1-4). IEEE.
12. Wang, R., Hong, W., & Yang, N. (2009, August). The research on footprint recognition method based
on wavelet and fuzzy neural network. In 2009 Ninth International Conference on Hybrid Intelligent
Systems (Vol. 3, pp. 428-432). IEEE.
13. Kumar, V. D. A., & Ramakrishnan, M. (2013). A comparative study of fuzzy evolutionary techniques for
footprint recognition and performance improvement using wavelet–based fuzzy neural
network. International journal of computer applications in technology, 48(2), 95-105.
14. Basheer, S., Nagwanshi, K. K., Bhatia, S., Dubey, S., & Sinha, G. R. (2021). FESD: an approach for
biometric human footprint matching using fuzzy ensemble learning. IEEE Access, 9, 26641-26663.
15. Vera-Rodriguez, R., Mason, J. S., Fierrez, J., & Ortega-Garcia, J. (2012). Comparative analysis and
fusion of spatiotemporal information for footstep recognition. IEEE transactions on pattern analysis and
machine intelligence, 35(4), 823-834
16. Vera-Rodriguez, R., Fierrez, J., Mason, J. S., &Orteua-Garcia, J. (2013, June). A novel approach of
gait recognition through fusion with footstep information. In 2013 International Conference on
Biometrics (ICB) (pp. 1-6). IEEE.
17. King, R. R., & Xiaopeng, W. (2013). Study of biometric identification method based on naked
footprint. International Journal of Science and Engineering, 5(2), 29-35.
18. Edwards, M., & Xie, X. (2014, October). Footstep pressure signal analysis for human identification.
In 2014 7th International Conference on Biomedical Engineering and Informatics (pp. 307-312). IEEE.
19. Moorthy, T.N., Mostapa, A.M.B., Boominathan, R. and Raman, N. Stature estimation from footprint
measurements in Indian Tamils by regression analysis. Egyptian Journal of Forensic Sciences, 4(1), 7-
16 (2014)
20. Moorthy, T.N. and Sulaiman, S.F.B. Individualizing characteristics of footprints in Malaysian Malays for
person identification from a forensic perspective. Egyptian Journal of Forensic Sciences, 5(1),13-22.(
2015)
21. Devadoss, A.K.V., Malaishamy, R., Subramaniam, M. and DEVADOSS, A.K.V. Performance
improvement using an automation system for recognition of multiple parametric features based on
human footprint. Kuwait J. of Science, 42(1). 2015
22. Kumar, V.A., Malathi, S., Kumar, V.A. and Kannan, P. Performance improvement using an automation
system for segmentation of multiple parametric features based on human footprint. Journal of Electrical
Engineering & Technology, 10(4),1815-1821.(2015)
23. Khokher, R., Singh, R.C. and Kumar, R., January. Footprint recognition with principal component
analysis and independent component analysis. In Macromolecular Symposia.347 (1), 16-26. (2015)
24. Costea, M., Sarghie, B., Mihai, A., & Rezus, E. (2017). Classification of the elderly foot types based on
plantar footprints. Procedia engineering, 181, 36-43.
25. Costilla-Reyes, O., Vera-Rodriguez, R., Scully, P., &Ozanyan, K. B. (2016, October). Spatial footstep
recognition by convolutional neural networks for biometric applications. In 2016 IEEE SENSORS (pp.
1-3). IEEE.
26. Nagwanshi, K.K. and Dubey, S. Statistical feature analysis of human footprint for personal identification
using BigML and IBM Watson Analytics. Arabian Journal for Science and Engineering, 43(6), 2703-
2712. (2018)
27. Nagwanshi, K.K.Cyber Forensic Review of Human Footprint and Gait-based System for Personal
Identification in Crime Scene Investigation. (2018).doi:10.20944/preprints201804.0072.v1
28. Abuqadumah, M. M., Ali, M. A., AbdAlmisreb, A., &Durakovic, B. Deep transfer learning for human
identification based on footprint: A comparative study. Periodicals of Engineering and Natural
Sciences, 7(3) 1300-1307. (2019)
29. Ibrahim, Y. I., & Alhamdani, I. M. A hyprid technique for human footprint recognition. International
Journal of Electrical & Computer Engineering 9, 2088-8708. (2019)
30. Kushwaha, R., Singal, G., & Nain, N... A texture feature based approach for person verification using
footprint bio-metric. Artificial Intelligence Review, 1-31. (2020)
31. Khokher, R., & Singh, R. C. Footprint identification: Review of an emerging biometric trait. In
Macromolecular (2021, June)
32. Kumar, A., & Shekhar, S. Personal identification using multi biometrics rank-level fusion. IEEE Trans.
Syst., Man, Cybern. C, Appl. Rev., vol. 41,no. 5, pp. 743-752.(Sep. 2011)
33. Pataky, T. C., Mu, T., Bosch, K., Rosenbaum, D., & Goulermas, J. Y. Gait recognition: highly unique
dynamic plantar pressure patterns among 104 individuals. Journal of the Royal Society Interface, 9(69),
790-800. (2012)
34. Gupta, A., & Raj, D.Novel distance metric for touch less footprint based identification technique. Int. J.
Innov. Technol. Exploring Eng., 9(3), 1011-1016. (2020) doi: 10.35940/ijitee.C7967.019320.
35. Gupta, A., & Raj, D. Comparative Analysis of Texture-Based Algorithms LBP, LPQ, SIFT, and SURF
Using Touchless Footprints. In International Conference on Artificial Intelligence and Sustainable
Engineering (pp. 423-439). Springer, Singapore. (2022)
36. Anshu, G., & Deepa, R. Footprint Recognition Using Invariant Feature Extraction Techniques. Far East
Journal of Electronics and Communications 24(2) (2021), 81-108. doi: 10.17654/EC024020081
37. Nagwanshi K. K. and S. Dubey. Dataset: Biometric 220 × 6 human footprint. IEEE Dataport, 2019.doi:
10.21227/7gmx-jq63.
38. R. Kumar. (2019). Footprint Image Database: Dataset of Footprint Images of 21 Individuals. [Online].
Available: https://github.com/goodrahstar/footprint-database
39. https://towardsdatascience.com/principal-component-analysis-for-dimensionality-reduction
115a3d157bad
40. Theodoridis, S., & Koutroumbas, K. Supervised learning: The epilogue. Pattern Recognition, 4th
edition, Boston, MA, USA: Academic pp. 567-594 (2009).doi:10.1016/B978-1-59749-272-0.50012-8
41. Chen, W. H., Hsu, S. H., & Shen, H. P. (2005). Application of SVM and ANN for intrusion
detection. Computers & Operations Research, 32(10), 2617-2634.
42. Zupan, J. (1994). Introduction to artificial neural network (ANN) methods: what they are and how to use
them. Acta Chimica Slovenica, 41, 327-327.
43. Mahesh, B. (2020). Machine learning algorithms-a review. International Journal of Science and
Research (IJSR).[Internet], 9, 381-386.