23/11/2021, 19:02 Subspace Analysis Technique on Features Images in Leaf Species Recognition | IEEE Conference Publicatio…

IEEE.org IEEE Xplore IEEE-SA IEEE Spectrum More Sites SUBSCRIBE SUBSCRIBECart Create Account Personal Sign In 

 Browse  My Settings  Help  Institutional Sign In

Institutional Sign In

All  

ADVANCED SEARCH

Back to Results
Conferences > 2021 IEEE Madras Section Conf... 

Subspace Analysis Technique on Features Images in Leaf

Species Recognition
Publisher: IEEE Cite This  PDF

<< Results

V. Vinay Kumar ; K.V. Ranga Rao ; V. Sridhar All Authors


2 More Like This
Full Alerts
Text Views
Texture feature extraction using 2-D Gabor
Manage Content Filters
Alerts 2012 International Symposium on Computer
Applications and Industrial Electronics
Add to Citation
(ISCAIE)
Alerts Published: 2012

Feature Extraction for Face Recognition

Based on Gabor Filters and Two-Dimensional
Abstract Locality Preserving Projections

Downl 2009 Fifth International Conference on
Document Intelligent Information Hiding and Multimedia
PDF
Sections Signal Processing
Published: 2009
I. Introduction Abstract:The aim of the proposed work is to classify the leaf species from an
image of a leaf with computer vision. For this the input image will be processed Show More
II. Literature with image processin... View more
Survey

III. System
 Metadata
Methdoology Abstract:
The aim of the proposed work is to classify the leaf species from an image of a
IV. Experiments, leaf with computer vision. For this the input image will be processed with image
Results and
processing algorithms, analyzed to extract features and then classified based
Discussion
on the extracted features with the help of pattern analysis and recognition
V. Discussion techniques. We present a robust leaf species recognition system based on
Gabor wavelet. Multiple Gabor wavelets with different scale and orientations are
Show Full Outline
 used to extract features. To have robust discriminative features which will also
reduce the computational complexity and memory cost, we have used Gabor
wavelet with selected scale and orientations only. For this selection, we
Authors
performed number of experiments taking a varied number of scales and
orientations. The feature vector so obtained has high dimensionality. The
Figures
feature dimension is reduced by down sampling & subspace analysis. The
kernel based subspace analysis is used to extract the nonlinear leaf variations.
References
The performance of the proposed system is extensively tested using the Folio
database. The results show a considerable saving in terms of computational
Keywords
cost and improvement in recognition rate compared to the contemporary related
work.
Metrics
IEEE websites place cookies on your device to give you the best user experience. By using our websites,
Accept & Close
you
Moreagree to the placement
Like This of these
Published in: cookies.
2021 IEEE ToSection
Madras learn more, read our
Conference Privacy Policy.
(MASCON)

https://ieeexplore.ieee.org/document/9563482/authors#authors 1/3
23/11/2021, 19:02 Subspace Analysis Technique on Features Images in Leaf Species Recognition | IEEE Conference Publicatio…

Date of Conference: 27-28 Aug. 2021 DOI:

10.1109/MASCON51689.2021.9563482
Date Added to IEEE Xplore: 19
October 2021 Publisher: IEEE

 ISBN Information: Conference Location: Chennai, India

V. Vinay Kumar
Anurag University, Ghatkesar, India

K.V. Ranga Rao

Neil Gogte Institute of Technology, Hyderabad, India

V. Sridhar
Vidya jyothi Institute of Technology, Aziz Nagar, India

Contents

I. Introduction
The main concern of this work is in extracting the effective
features having strong discrimination ability for a robust leaf
species recognition system. We propose a leaf species
recognition system based on the use of Gabor wavelets for
feature extraction. The feature extraction by Gabor wavelets is
followed by down sampling to eliminate the redundant data but
keeping the information intact. It is followed by kernel linear
discriminant analysis. Due to the application of subspace
Sign in to Continue Reading
analysis technique, the dimensions of the feature vectors are
further reduced. To make the separation between different leaf
species possible, the nonlinear kernel methods are applied to
map the data into higher order space for the separation of
different classes before the classification is done. This approach
has advantage compared to the linear subspace techniques like
principal component analysis (PCA) & linear discriminant
analysis (LDA) in terms of data separation.

Authors 
V. Vinay Kumar
Anurag University, Ghatkesar, India

K.V. Ranga Rao

Neil Gogte Institute of Technology, Hyderabad, India

V. Sridhar
Vidya jyothi Institute of Technology, Aziz Nagar, India

Figures 
References 
Keywords 
Metrics 

IEEE Personal Account Purchase Details Profile Information Need Help? Follow
IEEE websites place cookies on your device to give you the best user experience. By using our websites,
Accept & Close
you agree
CHANGE to the placement of these
USERNAME/PASSWORD cookies.
PAYMENT OPTIONSTo learn more, read our Privacy
COMMUNICATIONS Policy.
PREFERENCES US & CANADA: +1 800 678 4333   

https://ieeexplore.ieee.org/document/9563482/authors#authors 2/3
23/11/2021, 19:02 Subspace Analysis Technique on Features Images in Leaf Species Recognition | IEEE Conference Publicatio…

VIEW PURCHASED DOCUMENTS PROFESSION AND EDUCATION WORLDWIDE: +1 732 981 0060

TECHNICAL INTERESTS CONTACT & SUPPORT

About IEEE Xplore | Contact Us | Help | Accessibility | Terms of Use | Nondiscrimination Policy | IEEE Ethics Reporting  | Sitemap | Privacy & Opting Out of Cookies
A not-for-profit organization, IEEE is the world's largest technical professional organization dedicated to advancing technology for the benefit of humanity.

© Copyright 2021 IEEE - All rights reserved. Use of this web site signifies your agreement to the terms and conditions.

IEEE Account Purchase Details Profile Information Need Help?

» Change Username/Password » Payment Options » Communications Preferences » US & Canada: +1 800 678 4333
» Update Address » Order History » Profession and Education » Worldwide: +1 732 981 0060
» View Purchased Documents » Technical Interests » Contact & Support

About IEEE Xplore | Contact Us | Help | Accessibility | Terms of Use | Nondiscrimination Policy | Sitemap | Privacy & Opting Out of Cookies

A not-for-profit organization, IEEE is the world's largest technical professional organization dedicated to advancing technology for the benefit of humanity.
© Copyright 2021 IEEE - All rights reserved. Use of this web site signifies your agreement to the terms and conditions.

IEEE websites place cookies on your device to give you the best user experience. By using our websites,
Accept & Close
you agree to the placement of these cookies. To learn more, read our Privacy Policy.

https://ieeexplore.ieee.org/document/9563482/authors#authors 3/3
 Subspace Analysis Technique on Features Images
in Leaf Species Recognition

V. Vinay Kumar K.V. Ranga Rao V. Sridhar

Associate Professor Professor & HOD Associate Professor
Department of Electronics and Department of Computer Science and Department of Electronics and
Communication Engineering, Anurag Engineering, Neil Gogte Institute of Communication Engineering, Anurag
University, Venkatapur, Ghatkesar, India Technology, Hyderabad, India University, Venkatapur, Ghatkesar, India

Abstract—The aim of the proposed work is to classify the leaf We perform an empirical analysis in the selection of Gabor
species from an image of a leaf with computer vision. For this the wavelets. And based on the results some conclusions
input image will be processed with image processing algorithms, regarding the application of Gabor wavelet in leaf species
analyzed to extract features and then classified based on the recognition applications are drawn. We started our
extracted features with the help of pattern analysis and experimentation with a set of 40 Gabor wavelet filters with 5
recognition techniques. We present a robust leaf species scales and 8 orientations and then go on reducing the number
recognition system based on Gabor wavelet. Multiple Gabor of filters used. Based on the results obtained in these
wavelets with different scale and orientations are used to extract experimentation, we have fixed minimum conditions and
features. To have robust discriminative features which will also
requirements for further experimentation where we started
reduce the computational complexity and memory cost, we have
used Gabor wavelet with selected scale and orientations only. For
with single Gabor wavelet and go on increasing with scale and
this selection, we performed number of experiments taking a orientation. The nonlinear kernel LDA is used on the Gabor
varied number of scales and orientations. The feature vector so extracted features. These dimensionality reduced feature
obtained has high dimensionality. The feature dimension is vector are then evaluated with different similarity measure.
reduced by down sampling & subspace analysis. The kernel Finally for classification we used nearest neighbour classifier.
based subspace analysis is used to extract the nonlinear leaf The contributions made by this work are: First, a detailed
variations. The performance of the proposed system is analysis of Gabor wavelets from the point of view of
extensively tested using the Folio database. The results show a
decomposition of image into multiple oriented and spatial
considerable saving in terms of computational cost and
improvement in recognition rate compared to the contemporary
frequency channels and second, the effective discriminative
related work. feature selection based on the properties of the textures present
in the leaf image. The Organization of the paper is such that
Keywords—Local Discriminanat Analysis, Gabor Wavelets, followed to the introduction, next section is on literature
Feature Vector, Dimensionality Reduction survey. After which we are providing the methodology used in
building the system and then in next section we presented the
results obtained, followed by conclusions
I. INTRODUCTION
The main concern of this work is in extracting the effective
II. LITERATURE SURVEY
features having strong discrimination ability for a robust leaf
species recognition system. We propose a leaf species A. Leaf
recognition system based on the use of Gabor wavelets for
feature extraction. The feature extraction by Gabor wavelets is The leaf features can be divided into shape, interior
followed by down sampling to eliminate the redundant data structure of veins, the texture and colour. Leaf has a complex
but keeping the information intact. It is followed by kernel structure. The area occupied by the leaf contains midrib and
linear discriminant analysis. Due to the application of braches and sub branches. These are erected producing non
subspace analysis technique, the dimensions of the feature monotonic illumination variations and shadowing effects.
vectors are further reduced. To make the separation between These branches of the leaf image are with varied intensities
different leaf species possible, the nonlinear kernel methods indicating the presence of edges which corresponds to higher
are applied to map the data into higher order space for the frequency components. The rest of the part of leaf may be
separation of different classes before the classification is done. with uniform intensities or with some textures formed over it.
This approach has advantage compared to the linear subspace To extract the features from such complex structure is a
techniques like principal component analysis (PCA) & linear dedicated process involving its complete analysis.
discriminant analysis (LDA) in terms of data separation.
The leaf shapes are in different varieties and it is the most
Identify applicable sponsor/s here. If no sponsors, delete this text box useful feature in classification of the leaf species. Further the
(sponsors).
features based on shape are easier to extract. Such works on features, color moments and gray-level co-occurrence matrix
the leaf shape are found in [1], [2], [3], and [4]. Some leaves (GLCM). To implement the identifications systems, two
have complex structures and have more than one lobe which approaches has been investigated. First approach used a
may get overlapped and make it difficult to identify the leaf distance measure and the second used Probabilistic Neural
species, i.e. classification. Networks (PNN). The results show that Zernike Moments
Some researchers have made use of interior structure of have a prospect as features in leaf identification systems when
leaf which contain the main vein and from it number of they are combined with other features. The accuracy using
smaller veins are originated and run towards the peripheral Euclidean distance metrics is 94.69%
side of the leaf. If the leaf has some part occluded by other Wang et al.[6] describes a novel counting-based leaf
leaf, it becomes difficult to extract features [2]. recognition method, which combines leaf contours, textures
Identifying the colour of the leaves is fairly simple and it and veins in leaf images.
covers a range of different shades of green colour. But use this Here the local line responses are obtained by convolving
colour information only to classify the leaf species will not elliptical half Gabor with the grayscale leaf images. From the
prove better. This is due to the reason that the leaf changes its
local line responses, maximum gap local line direction
colour based on its location on the tree and also based on the
patterns are extracted and normalized indirection by cyclically
season. The texture of the leaf from front side and from back
side. It covers entire surface of the leaf including the inner right shifting these patterns. The histogram of the normalized
structure of main vein and other veins. The pre-processing patterns is calculated and regarded as the counting-based local
methods involved in extracting the above mentioned features structure descriptor, and support vector machine is utilized as
works on different type of image like RGB colour images, the classifier.
Grey images and binary images. Experimental results on three frequently used leaf
Sample points are taken on the leaf image to extract its databases-Swedish, Flavia and ICL show classification
geometrical features. From these sample points various accuracy of 98.4%, 97.83% and 96.76% respectively.
features like area, eccentricity, perimeter, aspect ratio etc. can Ji-Xiang Du et al. [7] used digital morphological based
be extracted. Feature vectors/descriptors are formed using feature extraction along with moving median centres (MMC)
these features and used for the classification of leaf species. hyper sphere classifier is used which gives recognition rate up
The curvature and the total curved length can be determined to 93%.
by connecting the sample points across the contour. The shape
of this contour is different for different leaves. Yaonan Zhang et al. [8] use bag of features (BOF) for
classification. Dual-output pulse-coupled neural network
There are different methods in the literature to find the leaf (DPCNN) is used texture features for image segmentation. bag
species from the image of leaf. The methods used to extract of features and DPCNN (BOF_DP) and LDA (linear
the features of leaf image are classified as follows: discriminant analysis) algorithms are, respectively, employed
 Template matching for textual feature extraction and reducing the feature
 Fourier transform dimensionality. Linear Support vector machine (SVM) is used
 Wavelet transform for classifier. Proposed method gives an accuracy of 98.53 on
Flavia database.
 Subspace analysis
Stephen Gang Wu et al. [9] employ Probabilistic Neural
Combining the Gabor wavelet transform with subspace
Network (PNN) for automated leaf recognition for plant
analysis technique is the new area of research in leaf species
classification. 12 leaf features are extracted and orthogonal zed
identification system.
into 5 principal variables. These variables consist the input
In classification the various methods used are given below: vector of the PNN. The method is applied on the leaves of 32
kinds of plants. The obtained accuracy is greater than 90%.
 Classification Tree
 Artificial neural networks Neeraj Kumar et al. [10] describe the first mobile app for
 Nearest neighbour classifier identifying plant species, the system is called Leafsnap. An
additional provision of discarding non-leaf images is
Minggang Du et al. [2] discuss the shortages of LDA, then available. The Curvature based leaf’s features (leaf’s contour)
classify the plant by using PCA+LDA. SVM is used for are used by taking multiple scales. The dataset contains 184
classification. The experiments show that it is effective and trees in the North-eastern United States. Nearest neighbour
feasible for plant classification. The identification rate is 98% classifier is used.
and response time is 13 seconds. A similar work found by
Pedro F. Silva et al. [4] used principal component analysis Jyotismita Chaki et al. [11] proposed leaf species
(PCA). identification using two different shape modeling techniques, -
Moments-Invariant (M-I) model and CentroidRadii (C-R)
Abdul Kadir et al. [5] used Zernike moments to build model. For classification neural network classifier is used. The
foliage plant identification systems. In this case, Zernike data set consists of 180 images from three classes. The
moments were combined with other features: geometric accuracies obtained are ranging from 90%-100%.
 Ayan Chaudhury et al. [12] used contour based Gaussian function in the Gabor wavelet in x and y
classification using B-spline. The important contribution is in direction respectively. This sharpness value is directionally
classifying occluded plant leaves. Obtained current matches proportional to the central frequency in the overall Gabor
up to 74.8%. filter and inversely proportional to the width of the Gaussian.
Hence, if the width of the Gaussian is more, represents that
Based on the above survey, and from the theoretical
sharpness is less, the slope of the curve is small. Whereas if
analysis of any statistical recognition system given by
the Gaussian width is small, the curve is steeper and hence
A.K.Jain et al.[13], the textures in the image will play
the sharpness is more values is considered as “frequency
important role in the identification process. To extract these
signature” of the pixel. Hence we are using clustering in an
features using Gabor filters as given for the face recognition
N-dimensional space (N = scale x orientation) for getting the
system by Li Bai et al. [14], with multiple channel as given by
texture segmentation. These values are too oscillatory to get
Bovik et al. [15], in this paper we proposed subspace analysis
good local features. To alleviate this problem around each
technique using kernel LDA on Gabor featured leaf images.
pixel we consider a small local window and inside that
The theoretical aspects of Gabor wavelets and its advantage
window we compute the energy of the image. This energy
are well analysed by Maria and Pedro in their book [16].
represents the local energy of the image in the respective
frequency band as stated by Parseval’s theorem. These local
III. SYSTEM METHDOOLOGY energies are considered as ‘frequency signatures’ of the pixels
and are used for performing pixel by pixel classification to
A. Spatio-Frequency Representation get the image segmentation [15].
To analyse the frequency contents of an image at particular
location in a 2D space, it is necessary to simultaneously C. Multiple Channel Filter Model Using Gabor Wavelet
represent the image in space and frequency form. To answer Multiple channel filter refers to the filter bank with
this problem Gabor has extended the Gaussian function by narrowband filters all having similar structure separated by a
multiplying it with the complex sinusoidal signal whose fast narrow band covering the entire frequency range of the input
varying frequency can extract the high frequency contents in
signal or image. The Fig 1 shows the multiple channel filter
an image effectively [14].
model using Gabor wavelet for 5 different scales and 8
B. The Uncertainty Principle  = a represents the ratio of central frequency and b
The uncertainty principle in signal processing tells that the orientations. For each of the narrow band filter, its real part
product of the spectral bandwidth and the time duration of a sharpness in the x-directional and  = f represents the ratio of
signal must be greater than or equal to ¼. For image central frequency and sharpness in the y direction.
processing applications the product of the spectral bandwidth
and the spatial spread of an image cannot be less than a certain x and y are the widths of the Gaussian function
minimum value [16]. representing standard deviation in x & y direction
respectively. ϕ is the orientation of the major axis of elliptical
The 2D Gabor wavelet in spatial domain (ψ(x, y)) and Gaussian and imaginary part shown separately. The magnitude
frequency domain Φ(u,υ) [14] is given by of the filter is also shown for different scales, as the
ψ(x,y) = f 2 exp(-(a2 x2 + b 2y 2)) .exp(j2πfx ) (1) orientation of the filter does not cause any change in the
magnitude.
different scale and orientation to produce number of output
intensity images representing the features in the narrow band D. Kernel Linear Discriminant Analysis
specified by the Gabor filter, using equation (5) In the subspace analysis principal component analysis
g(x,y) = f(x,y) * ψ(x,y) (2) 1
 (u, v)  e 
2

2 (u r  f )2
(5)
 b 2( vr )
2
The Gabor function with a set of orientations and radial (a
distances is termed as Gabor filter bank. Commonly 40 filters
with 8 orientations and 5 scales are used. The set of these Where u,v are the frequency axes in the frequency domain
representation of the Gabor wavelets and ur and vr are the
 r r r
xr   & yr   (3) components in terms of the polar co-ordinates.
x.cos  y.sin   x.sin  y.cos 
The important parameter in the Gabor wavelet are:
Where the spatial cartesian coordinates are represented by x i) Central frequency f ,
and y. xr & yr is giving the polar form representation. The
ii) Orientation θ and
Gabor filter is a product of Gaussian function with the
iii) Ratio between frequency and the sharpness of Gaussian
sinusoidal function. ‘f’ represents the central frequency of the
sinusoidal function represents the sharpness of the axis (λ, ς).
Usually keeping this ratio fixed for different values of
2 2f
a ,b  (4) scale and orientations we design different Gabor wavelets.
x y Here it is important to note that by varying the scale parameter
we are varying the central frequency. To keep the ratio fixed, where {wi | i= 1,2, …, m} is the set of generalized
we are varying the sharpness by the same factor as that of eigenvectors of SB and SW for the m largest generalized
frequency. Increase in central frequency results in decreasing eigenvalues {λi | i= 1,2, …, m}
the size of the Gaussian envelope. The choice of central
frequencies is based on the Nyquist sampling rate and could The projection of the variance of the image on this
not cross the value 0.5 as specified in the optimality criteria projection matrix is given by
[15]. Further it is made narrow for leaf recognition as given in
[14] to 0.25 cycles/sec. Most of the works in leaf species y = kx. W (10)
recognition follow from the strategies proposed in [2][8][9].
which will give the feature vector for the image.
To filter the image with a Gabor wavelet for extracting the
spatial features in particular frequency domain, input intensity E. Block Diagram of Proposed System
image f(x,y) is convolved with the Gabor filter ψ(x,y), with To identify the leaf species after capturing its image or
used. The PCA is preferably used in the literature for photograph following steps are performed. The image is pre-
dimensionality reduction. It provides features that capture the processed and then features are extracted from it. From this
main directions along which leaf images differ to represent the feature code vector is formed. This code vector is then used to
leaf images effectively, but fails to lower the within-class get the identity of the leaf based on the code vectors of images
scatter. As there is no separation in terms of the leaf images of stored in the database in the form of codebook or from the
the different leaves. Each leaf image is treated in the same trained classifier.
way irrespective of the class to which it belongs. This
drawback of PCA is overcame by LDA by using the class The proposed system uses features extracted using Gabor
membership information [2]. Both these technique are linear wavelet bank. The discriminative features are selected using
and not sufficient to extract the nonlinear information from the extensive empirical analysis for selection of the scale and
leaf image. To extract the nonlinear information, we are using orientation of the wavelet. This is followed by down-
sampling and kernel based subspace analysis to reduce the
kernel techniques [13], where the data is first raised to higher
feature dimensionality. In down sampling, the alternate rows
values for providing the separability between different classes.
and columns of the filtered image are deleted at a sampling
After this the linear subspace analysis technique LDA is rate of 64. The kernel based subspace analysis is used to
applied. extract the nonlinear face variations. A projection matrix for
the training images is obtained. During the testing phase, the
k(x, y) = exp(−( x2/σ𝑥 + y2/σ𝑦)) (6)
input leaf image is convolved with the same number of Gabor
The output obtained after the kernel function applied is then filters. These Gabor features are then down sampled &
processed using linear discriminant analysis to get the combined. These combined features are then projected on the
discriminant feature vector. The within class ‘S w’ and kernel subspace and compared with the training feature
between class ‘SB’ scatter matrices are given respectively by vectors using Mahalanobis Cosine distance measure to
equation (5) and equation (6) recognize an input leaf image.
The architecture for the system using Gabor wavelet &
Sw  i 1 (xi  )(xi  )T
N
(7) kernel based discriminant analysis is given in Fig 1. The
images used in the experimentation are first pre-processed
where xi represents images in a class and μ represents the redundancy is reduced by down sampling with a factor of 64
mean of that class. The scatter matrix ‘Sw’ represents the as given by Bowik et al. These down sampled Gabor features
mean scatter of the sample vectors of different classes around are then converted into feature vector image wise. All such
their mean vector and made of equal size after aligning them. feature vector are concatenated to give the leaf feature vector.
Gabor wavelet bank containing filters with different scale and Even after down sampling the leaf feature vector may have the
orientation is used to extract features, which may be redundancy & more dimensionality. To reduce this
redundant. The dimensionality further, these vectors are used to learn kernel
fisher space. The available distinct features are then used for
B  i 1 (m  i )(m  i )T classification using nearest neighbour classifier.
N
(8)
Gallery
where ‘μi’ represents the mean of each class and m represents Leaf
Pre- Feature
Encoding
Feature Vector
Processing extraction Database
the mean of all classes. The scatter matrix ‘SB’ represents the Image
scatter of the mean vector of different classes around the mean Probe
Pre- Feature
Leaf Encoding Comparison
vector of all classes. Processing extraction
Image
The projection matrix is formed using Matched
TSBw Leaf
|w |
w  arg max  [w1 , w 2 ,..., w m ] (9) Image
wSw Fig 1. Block diagram of the proposed system
 To identify the unknown leaf species, it is given as input to than 64, we will not lose much of the information. We used
the system. It passes through Gabor filter bank, down here a down sampling factor of 16 & 64.
sampling by 64 and finally projected to the learned kernel
fisher space to obtain the distinct feature vector. This is then iv) Kernel Subspace Analysis: In kernel subspace analysis
compared with the feature vectors of gallery images stored in LDA is performed on the data which is mapped to the high
the database. From all the obtained comparisons the smallest valued feature space for the purpose of easy differentiation
distance is considered to recognise the unknown leaf species. using equation (6). Using the given kernel function the down
sampled Gabor feature vectors are mapped to high valued
The functions of the system are broadly divided into space where the distinctness between the samples is
Preprocessing, Feature extraction, and Classification. increased. In selecting the kernel function there is a wide
i) Pre-processing: To remove the illumination effects choice including square function, cube function or any
polynomial function. Due to its characteristic nature the
adaptive histogram equalization and gamma correction is
Gaussian function is widely used as a kernel function. The
used. For cropping of leaf from the database image and further
value of the spread parameter (standard deviation of the
processing, the midrib of the leaf is used. Based on the midrib Gaussian function is taken 0.75 as mentioned in [18]. Later
coordinates the leaf images are aligned, normalized and finally the LDA is performed on this data to obtain the distinct
resized into 128x128 size. For resizing the image bilinear feature vectors with reduced dimension.
interpolation is used. In interpolation we are creating new
pixels with intensity levels within the range of a known pixel
values. The grid of required size, here 128x128 is created. The
intensity values are assigned to these pixels in grid from the
image pixels. Here the linear interpolation is performed in x
direction first and then in y direction, hence the name is
bilinear.
ii) Feature extraction: The input leaf image is high
dimensional, i.e., the number of leaf images per object in the
gallery set are far less as compared to the number of pixels
present in the image. The input image with size 128x128 is
said to have dimensionality 128x128, i.e. 16384. This
dimensionality is further increased in the multiples of the
number of wavelets used in the filter bank. After convolving
the image with Gabor filter bank with 5 scales and 8
orientations, the dimensionality is increased by 5x8 = 40 Fig 2. Leaf image and 5x8 Gabor wavelets
times. This dimensionality is reduced by down sampling and The dimension of the discriminant analysis are determined
discriminant analysis process as given in the next step. by the number of objects present. If there are C number of
Fig 2 shows the real part, imaginary part and magnitude of objects, the dimension of feature vector is C-1. The feature
all Gabor wavelets. Each row represents a different scale vectors so obtained for the gallery images are stored into
whereas each column represents different orientations. The memory.
input leaf image shown in Fig 2(a) is convolved with the Consider an image with size 128x128. If convolved with
Gabor filters in the bank to get Gabor feature set. The Gabor 5x8 filter bank, the total features obtained will be
feature set thus consists of convolution results of an input 128x128x5x8=655,360 features. Due to such high dimension
image with all of the 40 Gabor wavelets. It is shown in Fig 3. the computational cost is very high. To reduce the high
Our aim is to reduce the number of Gabor wavelets. Starting dimensionality and subsequent memory and computational
with all the filters we reduced the number of filters to 12. We cost down sampling and kernel subspace analysis is
used Gabor wavelets with 3 scales and 4 orientations to extract performed. With a down sampling rate of 64 Gabor features
the image features. Hence the dimensionality of the feature for for the example image can now be reduced to 128 x 128 x 5 x
a leaf image will be equal to 128x128x3x4. Which will be 8/64 = 10,240. When projected on the kernel subspace, for
then down sampled as given in the next step. database of 32 classes, the feature dimension will be 31. We
iii) Down sampling: If we use such dimension in the have taken 3x4 Gabor wavelets:
processing then not only the computation cost will increase Single scale is selected and for different no. of
but also the memory requirement to hold the data will increase orientations the leaf species recognition results are
and the process will consume more time. A solution for this is determined. This is performed on every scale. For two
suggested by [16] in their work carried on the representation different scales, the orientations with best results obtained in
of the image by the reduced number of Gabor features. The step (1) are tested. This will be carried with all 5 scales and
author has shown that the Gabor feature image contains best orientations obtained. Finally the optimal scales and
redundant data and even if we down sample it by a factor less optimal orientations are combined together to produce the
reduced Gabor filter bank.
 The total number of Gabor wavelet used are 12 with 4 measure. The performance of the system is evaluated by
orientation 30, 60, 120 and 150 degree and three scales with performing number of experiments on publicly available leaf
final values as 0.25, 0.111 & 0.077. Followed by the Gabor species databases- Folio. The parameter for checking the
feature extraction, Kernel LDA subspace analysis with performance of the system is the recognition rate and time
Gaussian kernel is used in the high dimensional feature taken.
space. For this the Gaussian kernel is used.
Recognition Rate = (Number of leaf species
recognized)/(Total leaf species in the test dataset) * 100
(11)
In all the experiments, we compared our proposed
approach of reduced number of Gabor filters with the
extensive set of Gabor filters.
Experimental comparisons on the Folio database: From the
available extensive set, the leaf images of 32 objects, with
10 images per object are taken. All the images were
normalized and cropped to 128 x 128 pixels. The data is used
as gallery sets and probe sets [15].

Fig 3. Gabor Filtered Images of Fig 2 (a)

v) Classification: For the input test image the similar

process as described in the above steps is carried out.
The image after pre- processing, is filtered with Gabor
wavelet bank, down sampled, mapped to high valued
space and then projected into kernel subspace to obtain
the leaf feature vector. The feature vector is compared
with the feature vectors stored in the database. A nearest Fig 4 Some images in the Folio database[18]
neighbour (NN) classifier is used as a similarity TP= FP=07 128 TP=1 FP=0 152
measure. FN TN=561 396 FN=0 TN=5 3968
128 3968 409 152 5624 4096
IV. EXPERIMENTS, RESULTS AND DISCUSSION Fig 5 The confusion matrix for leaf images in Folio database. Column1-
The number of filters in the reduced filter bank are complete 5x8 filter bank, Column2- 3x4 filter bank
selected based on the empirical analysis performed with
different scale and orientation. The Table 1 gives some of the The Matthews correlation coefficient (MCC) is
observations performed with different similarity measure. considered as a balanced measure in determining the binary
classification. It can also work effectively on classes with
different size. Its value lies between -1 to +1. +1 indicates
TABLE I. RESULTS OF EXPERIMENTS WITH DIFFERENT NUMBER OF FILTERS
perfect prediction. Hence more positive the value of MCC
Sl. Number Recognition Rate obtained with different better is the classification.
No. of filters MahCos Euclidean Cosine It is given by
1 1x4 17.50 72.50 70.83
2 2x4 54.17 81.67 86.67 TP * TN  FP * FN
MCC  (12)
3 2x8 20.00 70.83 74.17 (TP  FP)(TP  FN)(TN  FP)(TN  FN)
4 3x2 90.00 75.83 75.00
The MCC value obtained for the leaf images are 0.865,
5 3x4 93.33 89.17 88.33 0.853.
6 4x4 90.00 90.00 89.17
7 4x8 87.50 90.00 90.83 From the above experiments on leaf species recognition, it
is observed that: The orientation of Gabor filter is important in
8 5x4 90.83 90.83 90.83 extracting the discriminant leaf species features. The
9 5x8 93.33 91.67 92.50 computational and memory cost can be reduced considerably.

The highlighted part is showing the results obtain are equal From the Table 2, it is clear that instead of taking 40
Gabor wavelets, even with 12 Gabor wavelets (4x3), the
for Mahalanobis Cosine distance measure. Hence for rest of
results obtained are almost equal in terms of recognition rate,
the experimentation we preferred to use the same similarity whereas the time taken for execution is reduced.
 TABLE II.COMPARISON RESULTS OF FOLIO LEAF SPECIES DATABASE VI. CONCLUSION
S D Dataset1 Dataset Data
In this paper, the discriminative Gabor based leaf species
Fi RR/Time RR/Tim RR/
recognition is presented. Gabor wavelets are used to extract
2 8 95.0/36.0 68.0/36. 66.0 features from the leaf images. These features are obtained
3 4 95.0/25.0 67.5/23. 66.5 through the empirical analysis of different combination of
Confusion matrix and the parameters derived from it: We Gabor wavelet with different scale and orientation to check the
can convert our multi-class problem into binary classification distinctiveness of the features. Nonlinear kernel subspace
by labelling the outcomes either as positive (p) or negative (n). analysis technique is used to reduce the dimensionality and to
For this the object to be identified is taken as true value and all improve the discrimination ability of discriminative Gabor
other objects as imposter. For the given positive input, the features further. The experiments are performed on the
predicted output may be positive or negative. Same is the case popular Folio leaf species database.
for negative input. This work tested the possibilities of using different number
In the experiments carried out on the Folio database, we of Gabor filters with different combination of scale and
have 32 objects. The training images are 3 per object and each orientations. A reduced set of Gabor filter is proposed to
of the testing dataset contains 4 images per object. Hence the capture the discriminative features. A model of robust leaf
number of true images in the database are 32x4=128 and the species recognition system is developed based on the narrow-
total number of imposter images is (32-1) x128 = 3968. The band modelling of input leaf image, modelling the channel
total population is 4096. In the figure given below the first filters with different scale and orientation and finally a
column gives the confusion matrix obtained with 5x8 Gabor projection model to further reduce the dimensionality of the
filter bank and second column gives the confusion matrix Gabor feature descriptor. The experiments demonstrate that
obtained with 3x4 Gabor filter bank. the proposed discriminative & reduced set of Gabor filter
approach reduces the computation & memory cost. The
The various parameters obtained from the above confusion advantages of the proposed approach includes extraction and
matrix are listed in Table 5. The parameter accuracy is about selection is performed through empirical analysis, data
getting the true positive and true negative values from the total dimensionality can be reduced making the computation and
population. Sensitivity gives the recall or hit rate and it is true memory cost at the time of classification is reduced.
positive rate (TPR). Specificity is the selectivity of the
classifier and it gives true negative rate (TNR) by reducing the
REFERENCES
number of Gabor filters. The Gabor orientation in some
orientation may extract more discriminative features. The
performance may be improved by considering different scales
and orientation along with additional center frequencies. [1] Oluleye Babatunde, Leisa Armstrong, Dean Diepeveen and Jinsong
Leng, “A survey of computer-based vision systems for automatic
identification of plant species”, Journal of Agricultural Informatics.
V. DISCUSSION 2015 Vol. 6, No. 1, pp. 61-70
In the use of Folio dataset, out of 128 images, 120 images [2] Minggang Du and XianFeng Wang, “Linear Discriminant Analysis and
are identified correctly. Out of the eight images, not Its Application in Plant Classification”, 2011 Fourth International
Conference on Information and Computing, IEEE Computing Society,
recognized, there is either too variation in the size and shape pp. 548-551
of the leaf or the variation in the rotation. [3] K. K. Thyagharajan, I. Kiruba Raji, “A Review of Visual Descriptors
and Classification Techniques Used in Leaf Species Identification”,
If we consider the nature of leaf species images in the first Archives of Computational Methods in Engineering (2019) 26, pp. 933–
and second row they are totally different. This is due to the 960 (Review Paper: Good)
fact that we are not comparing them by using template or [4] Pedro F. Silva, Andr´e R. S. Marcal, and Rubim M. Almeida da Silva,
geometrical features. The Gabor feature images after down “Evaluation of features for leaf discrimination”
sampling are transferred to higher values using kernel function [5] Abdul Kadir, Lukito Edi Nugroho, Adhi Susanto, P. Insap Santosa,
and then discriminant features are obtained. “Experiments of Zernike Moments for Leaf Identification”, Journal of
Theoretical and Applied Information Technology, Vol. 41 No.1, 2012,
pp. 82-93
[6] Xuan Wang, Weikang Du, Fangxia Guo , and Simin Hu, “Leaf
Recognition Based on Elliptical Half Gabor and Maximum Gap Local
Line Direction Pattern”, IEEE Access, VOLUME 8, 2020, pp. 39175-
39183
[7] Ji-Xiang Du, Xiao-Feng Wang, Guo-Jun Zhang, “Leaf shape based plant
species identification”, Applied Mathematics and Computation 185
(2007) 883–893
[8] Yaonan Zhang, Jing Cui, Zhao bin Wang, Jianfang Kang and Yufang
Min, “Leaf Image Recognition Based on Bag of Features”, Appl. Sci.
Fig 6 Leaf species images not recognized correctly. 2020, 10, 5177
[9] Stephen Gang Wu, Forrest Sheng Bao, Eric You Xu, Yu-Xuan Wang,
Yi-Fan Chang and Qiao-Liang Xiang, “A Leaf Recognition Algorithm
for Plant Classification Using Probabilistic Neural Network”
[10] Neeraj Kumar, Peter N. Belhumeur, Arijit Biswas, David W. Jacobs, W.
John Kress, Ida Lopez, and Jo˜ao V. B. Soares, “Leafsnap: A Computer
Vision System for Automatic Plant Species Identification”
[11] Jyotismita Chaki and Ranjan Parikh, “Plant leaf recognition using shape-
based features and neural network classifiers”, International Journal of
Advanced Computer Science and Applications (IJACSA), Vol. 2, No.
10, 2011 pp. 41-47
[12] Ayan Chaudhury and John L. Barron, “Plant Species Identification from
Occluded Leaf Images”, IEEE/ACM Transactions On Computational
Biology and Bioinformatics, Vol. 17, No. 3, May/June 2020, pp. 1042-
1055.
[13] Anil K. Jain, Robert P.W. Duin, and Jianchang Mao, “Statistical Pattern
Recognition: A Review”, IEEE Transactions on Pattern Analysis and
Machine Intelligence, Vol. 22, No. 1, January 2000, pp. 4-37
[14] Linlin Shen & Li Bai, “A review on Gabor wavelets for Face
Recognition”, Pattern Analysis Application, vol. 9, 2006, pp. 273-292
[15] Alan Conrad Bovik,“Analysis of Multchannel Narrowband Filters for We suggest that you use a text box to insert a graphic
Image Texture Segmentation”, IEEE Transactions on Pattern Analysis (which is ideally a 300 dpi resolution TIFF or EPS file with
and Machine Intelligence, Vol. 12, No. 1, Jan 1990, pp. 55-73 all fonts embedded) because this method is somewhat more
[16] Maria Petrou and Pedro Garcia Sevilla, “Image Processing Dealing with stable than directly inserting a picture.
Texture”, Wiley publication, 1st Edition, 2006, pp. 81-380
[17] https://en.wikipedia.org/wiki/Receiver_operating_characterist ic Access To have non-visible rules on your frame, use the
date 16/03/2021 MSWord “Format” pull-down menu, select Text Box >
Colors and Lines to choose No Fill and No Line.