2023 3rd International Conference on Mobile Networks and Wireless Communications (ICMNWC)

Capturing the Essence: Exploring Cutting-Edge

Feature Extraction Methods in Digital Image
Processing
1st Prabu S 2nd Natarajan B 3rd Roshan George
School of Computer Science and School of Computer Science and Department of Computer Science and
2023 3rd International Conference on Mobile Networks and Wireless Communications (ICMNWC) | 979-8-3503-1702-2/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICMNWC60182.2023.10435774

Engineering Engineering Engineering (AIE)

VIT University VIT University Amrita School of Computing, Amrita
Chennai, India Chennai, India Vishwa Vidyapeetham
[email protected] [email protected] Chennai, India
[email protected]

4th S.Sujitha 5th P Divya Bharathi 6th S. Rajeswari

Department of Computer Science and Department of Artificial intelligence & Department of Computer Science and
Engineering Data science Engineering
Roever Engineering College Panimalar Engineering College K. Ramakrishnan College of Engineering
Tamilnadu, India Chennai, India Tiruchirappalli, Tamilnadu
[email protected] [email protected] [email protected]

Abstract—Feature Extraction (FE) plays a vital role in the Several research works [1-3] have been carried out to develop
fields of image processing, machine learning and pattern new techniques and methods to minimize processing time and
recognition. Binarization, resizing, thresholding and improve accuracy. Many studies and demonstrations have
normalization are the initial steps applied before extracting been conducted to discover the best FE techniques. There are
features from an image. FE methods are applied to images to a few worth mentioning methods are Zernike Moments, HOG
extract the most essential features that help in detection, (Histogram of Oriented Gradients), LBP (Local Binary
recognition and classification tasks. FE techniques are used in Patterns), SIFT (Scale-Invariant Feature Transform), Zoning,
numerous applications, such as object recognition, character SURF (Speeded-Up Robust Features), CCV (colour
recognition and pattern recognition. This research study
Coherence Vector), BRISK (Binary Robust Invariant Scalable
presents a comprehensive research study of different image FE
techniques, including SURF, BRISK, PCA and SIFT, which
Keypoints), Grid-based colour Moment and Gabor Filter. The
have been employed for various applications in recent times. standard process followed by any application is pre-
This research study explains each method’s fundamental processing, FE and classification or detection or recognition.
concepts and pros and cons and demonstrates in what scenario This survey focuses on different FE techniques used in various
which feature extraction technique will be better. This research applications for the past five years.
work also presents a summary table of FE techniques used in
various applications with their accuracy percentage. II. FEATURE EXTRACTION TECHNIQUES
This section presents a brief introduction to each technique
Keywords—Image Processing, Feature Extraction, Image with its applications.
Representation, Feature Selection, Edge Detection.
A. Zernike Moments (ZM)
I. INTRODUCTION 1) Introduction: The Zernike moment was first proposed
A feature is an initial point and main primitive for many by Zernike in 1934. It remains one of the most popular
computer vision algorithms and subsequent algorithms. alternatives for some of the problems such as information
Features can be classified into two categories: local features redundancy, noise resilience and reconstruction capability in
and global features. The local features are typically geometric, feature extraction. Zernike polynomials’ orthogonal property
and the global features are topological. The main features of
images are edges, corners, blobs and ridges. The edge can be helps extract the independent information in an image.
defined as “a symbol and reflection of discreteness of partial 2) Works using ZMs: A method was introduced to
image; it can be of any shape.” It indicates the end of one area classify both global and local damages resulting from various
and the start of another area. A corner is “The points with impacts on composite materials [2]. The ZM technique was
different edge directions and two dominant in a local employed as a shape descriptor to analyze damages caused by
neighbourhood of the point.” A blob refers to a visual tensile loads in these materials. The features derived from ZM
component with points in an image that is brighter or darker exhibited enhanced accuracy in identifying both global and
than the neighbouring points. A ridge is another feature that local damages compared to geometric features. A novel local
forms a stable elevated crest for some distance. feature descriptor was proposed based on an adaptive
FE plays a significant role in image processing systems. It neighbourhood and embedding ZM for image matching. This
segregates image parts based on some special properties. approach aimed to achieve improved rotation invariance. ZM
Separating some parts from images can be referred to as was also utilized in iris and face recognition and in a
streaming. FE is a unique form of dimensionality reduction in fingerprint recognition system to evaluate fingerprint image
pattern recognition. The primary goal of FE is to obtain the features. Furthermore, ZMs extracted features from CT and
most relevant information from the original data and express MRI medical images. In designing an OCR system, ZMs
that obtained information in a lower dimensionality space.

979-8-3503-1702-2/23/$31.00 ©2023 IEEE

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on February 24,2024 at 09:48:53 UTC from IEEE Xplore. Restrictions apply.
served as a feature extraction tool to capture the distinctive detection and pattern recognition approaches. The main
features of each character. Linear Discriminant Analysis advantage of this algorithm is that it characterizes and
(LDA) was subsequently employed as a recognition tool for distinguishes the surface textures. It also produced high
character recognition from the calculated feature maps. In the accuracy in the low variance texture.
context of Copy-Move Forgery Detection (CMFD), the 2) Works using LBP: A research study utilized LBP-
pyramid model and ZMs were integrated. This involved non- based ear recognition features under controlled and
uniform sampling of the image, followed by rescaling with unconstrained imaging conditions. In the realm of brain
different scale factors. Using the ZMs method facilitated the tumor classification, two different LBP-based FE techniques,
extraction of image rotation-invariant features for subsequent nLBP and aLBP, were introduced. The final image was
tasks. generated using LBP, aLBP, and nLBP FE, and various
classifiers such as KNN, ANN, and RF were employed for
B. Histogram of Oriented Gradients (HOG)
the classification task. A modified LBP FE technique,
1) Introduction: The HOG algorithm was introduced by Generalized Completed Local Binary Patterns (GCLBP), was
Dalal in 2005 to extract a set of features for object introduced for steel surface defect classification. This
recognition. It is the most popular algorithm for human and approach included two additional techniques: Improved
pedestrian detection in the early stages. The HOG descriptor Completed Noise-invariant Local Structure Patterns (ICNLP)
can categorize the object’s shape and appearance accurately. and Improved Completed Local Binary Patterns (ICLBP).
It counts the number of occurrences of gradient orientation in Furthermore, the LBP method was utilized for classifying and
a detection window. Fig. 1 shows the workflow of the HOG extracting edge pixels from HyperSpectral Images (HSI),
algorithm. capturing spatial kernel information from the superpixels [4].
In a research study on breast cancer classification [5], the
capability of LBP was evaluated, with a CNN based on LBP
modality outperforming a CNN based on grayscale modality
(96.32% vs. 93.24% accuracy). Additionally, a new automatic
face recognition system was proposed using LBP,
demonstrating the effectiveness of the LBP-based FE method
under various real-world conditions such as noise, pose
variation, and different illumination. In the context of gender
identification, a system was developed utilizing facial
Fig. 1. Block diagram of the HOG algorithm.
elements like the mouth, nose, eyes, ear, forehead, and chin.
Features were extracted using the LBP algorithm, and an SVM
2) Works using HOG: An application was developed for classifier was employed for the classification task.
human detection in videos involving three key tasks: human D. Scale-Invariant Feature Transform (SIFT)
counting, human detection, and histogram generation. The 1) Introduction: SIFT is the most popular feature
HOG method was utilized, yielding a high level of detection detection technique in computer vision. It is used in various
accuracy. Consequently, HOG was strongly recommended applications such as object recognition, 3D scene tracking or
for human detection tasks. Another method that combined detection and panorama stitching. Fig. 2 shows the series of
HOG and SVM was introduced to detect power quality steps performed by SIFT to extract features.
events. This method proved effective in capturing both single
and multiple power quality events. HOG requires less
processing time than existing techniques, mainly when
applied to simultaneous events.
In the domain of facial expression recognition, a system
was developed consisting of four blocks: face processing,
domain transformation, FE, and expression recognition.
Combining HOG features with the DWT domain, termed as
W HOG feature, was employed to extract facial expression
features from the input image. An image segmentation Fig. 2. Illustration of the SIFT algorithm.
technique was proposed [3] based on HOG features,
facilitating the recognition of basketball referee signals from 2) Works using SIFT: The SIFT algorithm was employed
videos. Using the HOG algorithm, a framework was also for palm pattern recognition, with its performance compared
devised for exposing video inter-frame forgery. Here, the to the ORB algorithm demonstrating superior results in
HOG obtained from each image was considered a fingerprint recognition. A model was proposed [6] for
discriminative feature. One of the main drawbacks of HOG is identifying Lauraceae, utilizing SIFT and LDA for feature
the increase in processing time as the final descriptor vector extraction and species classification within the Lauraceae.
becomes longer. The model effectively classified 11 genera and 39 species. In
C. Local Binary Patterns (LBP) the domain of human ear identification, a system consisting
of four modules, feature extraction, ear detection, ear
1) Introduction: The LBP method is popularly used for
recognition, and confirmation, was developed. The SIFT-
texture FE. The LBP operator converts an image into an array
based feature extraction achieved recognition accuracy of
or integral label image. This method is very efficient for face
92%, outperforming the Gabor sampling filter (82%) and

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on February 24,2024 at 09:48:53 UTC from IEEE Xplore. Restrictions apply.
Gabor mean feature extraction (65%). An offline Arabic algorithms for extracting key points from different regions in
handwritten recognition system was proposed to tackle the an image using box filters. The SURF relies on the
challenging task of identifying Arabic handwritten characters determinant of the Hessian matrix (matrix of second
[7]. SIFT was initially used to eliminate redundant data and derivatives) for FE or blob detection. To a certain extent, it
capture a high-level representation of the character image. also eradicates the invalid feature points outside the
Subsequently, the Fast Approximate Nearest Neighbors overlapped area. SURF uses wavelet responses to calculate
(FANN) algorithm was applied to these features, followed by orientation by applying adequate Gaussian weights in both
SVMs for Arabic and non-Arabic character classification. vertical and horizontal directions.
Furthermore, the SIFT algorithm was utilized for object 2) Works using SURF: A research study utilized the
matching. In contrast, an integrated approach combining the SURF FE algorithm to detect instances of copy-move
advantages of SIFT and SURF algorithms was developed for forgery. The proposed work involved three critical processes,
currency recognition, outperforming the standalone methods encompassing feature extraction, description, and matching,
of SIFT and SURF. to derive a high-level representation of features from the input
image. Subsequently, an SVM was employed to identify
E. Zoning
copy-move forgery images.
1) Introduction: The zoning method is the most efficient Experiments were conducted to analyze the performance
and popular statistical-based FE technique that extracts the of SURF and SIFT algorithms in face recognition [13].
major information with low complexity and high speed. It is Similarly, in the domain of anomaly detection, both SURF and
mainly used for text recognition. It can be categorized into SIFT approaches were utilized, with the former demonstrating
two types: static and adaptive. Static topologies are based on superior performance [14]. In localizing an autonomous robot
a simple grid. It can be designed without a priori information without external tools, a proposed method employed three
of feature distribution in the pattern class. Adaptive zoning algorithms: SURF, ICP-BP, and EMD. Among these, SURF
exhibited superior performance compared to the other
topologies are the result of the optimization procedure for
methods.
zoning design.
2) Works using Zoning: Javanese characters are known Additionally, a system was developed to enhance the
for their diverse forms, presenting a challenge in performance of eye-gaze detection [15]. The SURF descriptor
understanding and memorization. A new Optimal Character was utilized to extract robust points from iris images,
Recognition (OCR) algorithm was proposed to facilitate generating a vector array of 64 dimensions for each iris image.
The system’s performance was evaluated using three
comprehension of these characters. This method adapts the
benchmark datasets, and the SURF FE technique greatly
Zoning FE technique to segment the character image into supported the proposed framework to improve its overall
multiple zones, with character identification based on the performance.
number of black pixels in each zone. Similarly, an OCR
technique was developed [8] for recognizing handwritten G. Colour Coherence Vector (CCV)
Pashto characters. A new dataset comprising 102 images of 1) Introduction: Colour coherence is defined as the
48 different Pashto characters and 4488 instances was degree to which image pixels of that colour are members of a
introduced. The approach involves using a Zoning algorithm large similarly colour region; these colour regions are
for FE, followed by Neural Network (NN) and KNN referred to as the coherent region. The coherence measure
classifiers for character classification. The KNN classifier categorizes pixels into any one of the two types: coherent or
was found to outperform the NN in this context. Zoning was incoherent. Incoherent pixels are not sizable contiguous
also employed for leaf disease identification [9], with the regions. Coherent pixels are sizable contiguous regions. The
performance of the Zoning technique evaluated using coherent vector is used to classify the colour in an image.
different classifiers such as SVM, K-NN, and ANN. CCV prevent coherent pixels in an image from matching
In the domain of number plate recognition, a technique incoherent pixels in another.
was proposed [10] that involved various phases, including pre- 2) Works using CCV: A machine learning framework was
processing, FE, segmentation, and detection and recognition. proposed to classify aerial images. This framework employed
The FE step combined zoning and statistical features to extract LBP and CCV approaches to acquire texture and colour
text from the number plate, significantly improving information. The performance of the fused algorithm was
recognition accuracy. Unified features derived from Zoning
and HoG features [11] were utilized to recognize handwritten, found to be slightly better than that of the individual methods.
printed, and scene text. The unified framework’s performance A fruit disease identification and classification system was
was assessed using Chars74K and EMNIST datasets. designed and developed in another research study [16]. The
Furthermore, a handwritten math symbol recognition model system extracted three main features—colour, texture, and
was developed [12], leveraging a hybrid FE technique that shape—using CCV, LBP, and ZM techniques. This approach
amalgamates the beneficial properties of the wavelet and achieved an accuracy of 96.29%, surpassing the performance
zoning approaches. The hybrid approach demonstrated of the state-of-the-art method (92.31%). Furthermore, CCV
superior performance compared to standalone methods. was used to identify leaf images’ chlorotic regions [17]. CCV
F. Speeded-Up Robust Features (SURF) demonstrated superior capabilities in extracting colour
features from specific regions compared to other colour
1) Introduction: The SURF algorithm is widely used in
histogram-based algorithms.
tracking, object detection, image registration and 3D
classification. It remains one of the fastest and most robust

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on February 24,2024 at 09:48:53 UTC from IEEE Xplore. Restrictions apply.
H. Binary Robust Invariant Scalable Keypoints (BRISK) and consumed limited computational power in various
1) Introduction: The BRISK algorithm was proposed by applications. It was particularly effective in accurately
handling image rotations and various geometric
Stefan Leuteneggerin in 2011. The typical BRISK algorithm
transformations, outperforming SURF and SIFT regarding
consists of three modules: Keypoint description, Keypoint feature identification.
detection and descriptor matching. The scale-space pyramid
is initially built, and stable extreme points of sub-pixel I. KAZE
precision are extracted using AGAST (Adaptive Corner 1) Introduction: The KAZE technique was proposed by
Detection Operator) from the continuous scale space. Then, Al-cantarilla in 2012. It is a nonlinear scale space. The three
the binary feature descriptor of the local image is established major steps in the KAZE feature extraction algorithm are: 1.
using a grayscale relationship of the random sample point Constructing a nonlinear scale-space pyramid of the original
pairs. Finally, Hamming distance is used for feature image. 2. Determine the key points by using Hessian
matching. determinants and multiscale derivatives. 3. Compute
2) Works using BRISK: A novel technique for copy-move orientation and build a descriptor vector for all key points.
forgery detection was proposed [18] based on image blobs KAZE method uses key points to recognize the regions of
and keypoints. This approach involved three key steps: first, sharp image changes. The additive Operator Splitting (AOS)
the extraction of blobs and BRISK features, which represent technique is the backbone for the KAZE to obtain constant
regions of interest in an image; second, the identification of nonlinear scale-spaces for any step size.
BRISK keypoints within these blob regions; and finally, the 2) Works using KAZE: A research study compared the
matching process, where BRISK keypoints from different performance of seven different feature extraction (FE)
blobs were compared to identify similarities. Similarly, other techniques in image 3D reconstruction. Among these
researchers utilized the BRISK FE algorithm in the context of techniques, KAZE demonstrated the least standard deviation
copy-move forgery identification. For instance, a system was and mean value, accurately capturing boundary information.
introduced to retrieve images from Arabic manuscripts, This facilitated subsequent steps in image matching and
incorporating BRISK and SURF for feature extraction and generating high-quality 3D reconstructed images. Similarly,
matching. BRISK was also employed for human face the performance of various FE techniques was compared for
recognition [19] to address challenges related to the global breast cancer classification in a separate research study [21].
feature generation process during feature extraction. Furthermore, an introduction was made to palmprint feature
Furthermore, in developing a fingerprint-based detection using KAZE, which exhibited the capability of
authentication system [20], the BRISK algorithm was utilized extracting more feature points from palmprint texture images,
to overcome challenges arising from scale, rotation, and essential for a palm-based authentication system. Another
illumination changes, thereby enhancing system palmprint authentication system was developed [22],
performance. combining KAZE and LBP feature-matching techniques. An
In the watermarking domain, a blind watermarking algorithm that utilized the KAZE FE method was proposed to
scheme for digital images was designed using the BRISK FE detect copy-paste forgery.
algorithm. Notably, BRISK was found to be faster than SURF

TABLE I. PERFORMANCE ANALYSIS OF DIFFERENT FE TECHNIQUES

Sl. No FE Techniques Application Recognition Accuracy
1 Zernike Moments (ZM) Iris and Face Recognition 87.2%
2 Scale-Invariant Feature Transform (SIFT) Offline Arabic Handwritten Recognition 90.1%
3 Histogram of Oriented Gradients (HOG) Facial Expression Recognition 84.3%
4 Local Binary Patterns (LBP) Steel Surface Defect Classification 85.5%
5 Zoning Leaf Disease Identification 90.2%
6 Speeded-Up Robust Features (SURF) Eye Gaze Detection 92.7%
7 Colour Coherence Vector (CCV) Fruit Disease Identification 88.4%
8 Binary Robust Invariant Scalable Keypoints (BRISK) Identify Copy-Move Forgery 94.6%
9 KAZE Breast Cancer Classification 89.8%
10 Oriented FAST and Rotated BRIEF (ORB) Video Stitching 97.0%

This system also incorporated the RANSAC algorithm to scaling and rotation. The ORB image-matching algorithm is
remove outliers and uncover forgery images. Notably, KAZE generally divided into three steps: extracting feature points,
required less computation time compared to SURF and SIFT. generating feature point descriptors and matching feature
J. Oriented FAST and Rotated BRIEF (ORB) points.
2) Works using KAZE: In an assessment of various
1) Introduction: The ORB method is an alternative to
descriptors, including SIFT, SURF, and ORB, in a 2D object
SIFT and SURF. ORB algorithm combines an enhanced
recognition problem [23], the combination of all three
FAST technique with a direction-normalized BRIEF
descriptors yielded more satisfactory results than using any
descriptor to locate key points. The FAST algorithm is
individual descriptor alone. Another research study combined
applied repeatedly to each layer in a scale pyramid to achieve
ORB and SIFT to extract hand-crafted features for an object
scale invariance. The corniness of detected FAST features is
recognition task. Additionally, a Currency Recognition
tested using a Harris measurement, and non-corner key points
System (CRS) was developed [24], employing the ORB and
are discarded. ORB method is invariant to image affine,

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on February 24,2024 at 09:48:53 UTC from IEEE Xplore. Restrictions apply.
SIFT algorithms to aid visually impaired individuals in recognition,” in Proc. of Iberoamerican congress on pattern
recognition, 2018, pp. 893-901.
identifying different currencies. Feature matching was
[13] S. Bilik, and K. Horak, “SIFT and SURF based feature extraction for
conducted using the Hamming Distance. A machine learning the anomaly detection,” arXiv preprint arXiv:2203.13068, pp. 1-7,
framework was proposed to identify objects in videos [25], 2022.
utilizing ORB to recognize candidate regions of an object. A [14] A. S. M. Yasin, M. M. Haque, M. N. Adnan, S. Rahnuma, A. Hossain,
multiclass SVM was then employed to verify and classify K. Naha, and F. Serratosa, “Localization of Autonomous Robot in an
Urban Area Based on SURF Feature Extraction of Images,”
these object regions. For video stitching, an algorithm was International Journal of Technology Diffusion (IJTD), vol. 11, no. 4,
designed that used ORB to identify the overlapping regions pp. 84-111, 2022.
between consecutive frames [26]. A pre-trained SVM [15] H. Farsi, R. Nasiripour, and S. Mohammadzadeh, “Eye gaze detection
facilitated the matching operation. Furthermore, a hybrid based on learning automata by using SURF descriptor,” Information
Systems & Telecommunication, vol. 6, no. 1, pp. 41-49, 2018.
feature extraction technique for facial expression recognition [16] S. R. N. M. Ayyub, and A. Manjramkar, “Fruit disease classification
was proposed, combining ORB and LBP to improve and identification using image processing,” in Proc. of 3rd International
computational efficiency during feature extraction. Detailed Conference on Computing Methodologies and Communication
(ICCMC), pp. 754-758, 2021.
information about the different feature extraction techniques
[17] A. M. Abdu, M. M. Mokji, and U. U. Sheikh, “Automatic vegetable
and their recognition accuracy can be found in Table I. disease identification approach using individual lesion features,”
Computers and Electronics in Agriculture, vol. 176, pp. 105660, 2020.
III. CONCLUSION
[18] P. Niyishaka, and C. Bhagvati, “Copy-move forgery detection using
Feature Extraction is essential in various fields, such as image blobs and BRISK feature,” Multimedia Tools and Applications,
image processing, computer vision, machine learning and vol. 79, no. 35, pp. 26045-26059, 2020.
pattern recognition. This research research study discusses an [19] K. M. Thi, “Face Recognition for Human Identification using BRISK
Feature and Normal Distribution Model,” International Journal of
overview of conventional feature extraction techniques and an Trend in Scientific Research and Development, vol. 3, no. 5, pp. 1139-
extensive review of recent enhancements, merits and demerits. 1143, 2019.
The results acquired by researchers by utilizing different FE [20] E. J. A. Taee, and Z. Abdulsamad, “A New Approach for Fingerprint
techniques are also organized together in a tabular format. Authentication in Biometric Systems Using BRISK Algorithm,”
This literature review aims to assist aspiring researchers in International Journal on Advanced Science Engineering Information
selecting suitable feature extraction techniques based on Technology, vol. 8, no. 5, pp. 1941-1947, 2022.
application, complexity and type of image. [21] C. Kaushal, and A. Singla, “Analysis of breast cancer for histological
dataset based on different feature extraction and classification
REFERENCES algorithms,” in Proc. of International conference on innovative
computing and communications, pp. 821-833, 2021.
[1] S. Prabu, and M. Sumathi, M, “Evaluating phishing website detection
on client side,” International Journal of Engineering and Advanced [22] N. A. A. Khalid, M. I. Ahmad, T. H. Mandeel, and M. N. M. Isa,
Technology, vol. 8, no. 6, pp. 18–22, 2019. “Palmprint features matching based on KAZE feature detection,” In
Journal of Physics: Conference Series, vol. 1878, no. 1, pp. 012055,
[2] H. Kaur and H. S. Pannu, “Zernike moments-based fingerprint 2021.
recognition using weighted-support vector machine,” Modern Physics
Letters, vol. 33, no. 21, pp. 1950245, 2019. [23] S. Gupta, M. Kumar, and A. Garg, “Improved object recognition results
using SIFT and ORB feature detector,” Multimedia Tools and
[3] R. Kapoor, R. Gupta, S. Jha, and R. Kumar, “Detection of power Applications, vol. 78, no. 23, pp. 34157-34171, 2019.
quality event using histogram of oriented gradients and support vector
machine,” Measurement, vol. 120, pp. 52-75, 2020. [24] A. Yousry, M. Taha, and M. M. Selim, “Currency Recognition System
for Blind people using ORB Algorithm,” International Arab Journal of
[4] P. Král, A. Vrba, and L. Lenc, “Enhanced local binary patterns for e-Technology, vol. 5, no. 1, pp. 34-40, 2023.
automatic face recognition,” in Proc. of International conference on
artificial intelligence and soft computing, pp. 27-36, 2019. [25] F. D. Adhinata, and A. Harjoko, “Object Searching on Video Using
ORB Descriptor and Support Vector Machine,” in Proc. of
[5] S. M. Osman, N. Noor, and S. Viriri, “Component-Based Gender International Conference on Computational Collective Intelligence, pp.
Identification Using Local Binary Patterns,” in Proc. of International 239-251, 2020.
Conference on Computational Collective Intelligence, pp. 307-315,
2019. [26] R. Yuan, M. Liu, M. Hui, Y. Zhao, L. Dong, L. Kong, and Z. Cai,
“Real-time video stitching based on ORB features and SVM,” In
[6] S. W. Hwang, K. Kobayashi, S. Zhai, and J. Sugiyama, “Automated Applications of Digital Image Processing XLI, vol. 10752, pp. 420-
identification of Lauraceae by scale-invariant feature transform,” 427, 2018.
Journal of Wood Science, vol. 64, no. 2, pp. 69-77, 2018.
[7] A. K. A. Hassan, B. S. Mahdi, and A. A. Mohammed, “Arabic
handwriting word recognition based on scale invariant feature
transform and support vector machine,” Iraqi Journal of Science, pp.
381-387, 2019.
[8] S. Khan, H. Ali, Z. Ullah, N. Minallah, S. Maqsood, and A. Hafeez,
“KNN and ANN-based recognition of handwritten Pashto letters using
zoning features,” arXiv preprint arXiv:1904.03391, 2022.
[9] A. Labellapansa, A. Yulianti, and A. Yuliani, “Segmentation of Palm
Oil Leaf Disease using Zoning Feature Extraction,” in Proc. of Second
International Conference on Science, Engineering and Technology, pp.
98-101, 2019.
[10] M. N. Tanurkar, M. V. Bhalerao, and A. A. Kadam, “Zoning based
Number Plate Recognition,” International Journal of Engineering
Research & Technology (IJERT), vol. 8, no. 7, pp. 748-753, 2019.
[11] M. Arun, and S. Arivazhagan, “A unified feature descriptor for generic
character recognition based on zoning and histogram of gradients,”
Neural Computing and Applications, pp. 1-12, 2022.
[12] C. R. Piña, J. S. Sánchez, R. M. V. Rosas, and J. A. H. Servín, “A hybrid
feature extraction method for offline handwritten math symbol

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on February 24,2024 at 09:48:53 UTC from IEEE Xplore. Restrictions apply.