International Journal of Environment, Agriculture and Biotechnology

Vol-9, Issue-4; Jul-Aug, 2024

Peer-Reviewed International Journal
Journal Home Page Available: https://ijeab.com/
Journal DOI: 10.22161/ijeab

Comparison of Adapted and Improved Feature Extraction

Techniques of Different Potatoes Types using Image
Processing
D. Ravindra Babu1*, R.C.Verma2, Navneet Kumar Agrawal3, Isha Suwalk4
1,2Ph.D. Scholar and Professor, Department of Processing and Food Engineering, Maharana Pratap University of Agriculture and
Technology, Udaipur, Rajasthan, India
3&4 Assistant Professor and Ph.D Scholar, Department of Electronics and Communication Engineering, College of Technology and

Engineering, Maharana Pratap University of Agriculture and Technology, Udaipur, Rajasthan, India
*Corresponding Author: [email protected]

Received: 22 Jun 2024; Received in revised form: 27 Jul 2024; Accepted: 03 Aug 2024; Available online: 13 Aug 2024
©2024 The Author(s). Published by Infogain Publication. This is an open-access article under the CC BY license
(https://creativecommons.org/licenses/by/4.0/).

Abstract— The characteristics of crack, rotten, sprout, skin peel and good potatoes non destructively with
gray level co-occurrence matrix properties (GLCMP), radon, gabor, local binary patterns (LBP) and
histogram of oriented gradients (HOG) with default parameters and values i.e. adapted method were
compared with improved method. Gabor feature length (16) of improved method was lower compared to
adapted method and improved method and it requires less time to plot gabor magnitude and spatial kernels
for all potato classes. Radon feature row vector size is same for both adapted and improved methods for all
potato classes but differ in column vector size. At theta value of 90° (improved method), the time taken to
plot radon transforms is lower compared to adapted method (using theta value 180°). Gray level co-
occurrence matrix properties (GLCMP) such as contrast, correlation, energy and homogeneity values were
compared to both adapted and improved methods for all potato types. Contrast values found lower in adapted
method for all potato classes compared to improved method. But remaining three properties found highest
in adapted method for all potato classes compared to improved method. The default values used in adapted
method of HOG feature vector length (26140) is higher compared to improved method (1330) for all types
of potato images. For crack and rotten potato images, an improved method required higher time to plot
visualization than adapted method, while for sprout, good and skin peel images, adapted method has more
visualization time. The LBP feature length in improved method was found higher (185) compared to adapted
method (59) for all potato classes. The mean time to plot squared errors in adapted and improved methods
for crack images were found to be 0.6378 s and 0.6305 s respectively, for rotten images 0.2098 s and 0.2622
s, for sprout images 0.1911 s and 0.2209 s, for skin peel images 0.2197 and 0.2197 s, for good images 0.2672
and 0.2565 s.
Keywords— Local Binary Patterns, Histogram of Oriented gradients, Radon Features, Gabor Features
and Grey Level Co-occurrence Matrix Properties

I. INTRODUCTION Image processing carries excellent verification of

The demand for image processing is increasing agricultural commodities (Unay et al., 2011). One of the
day by day by extracting preciuos data from images without most famous cultivations on earth, which is widely
touching the objects. An each image contains information consumed as raw or cooked, is potato (Solanum Tuberosum
and it is extracted either in the form of numerical or in L.). Farming of potatoes accounts for almost 80 per cent of
graphical fprm by different feature extraction techniques. countries. Following rice, wheat and maize, it is India’s

ISSN: 2456-1878 (Int. J. Environ. Agric. Biotech.)

https://dx.doi.org/10.22161/ijeab.94.22 161
Babu et al. Traditional Low-Efficient and Anti-Ecological Agriculture Must Be Replaced with Modern Cellular and
Hydroponic Agriculture

main crop. The need for potatoes for processed products is (2002) given that LBP features mainly depend on number
strong and not for fresh use. This increases the demand on of neighbors and radius. The general combinations used in
the market for high-quality potatoes (Rani and Prasoon, LBP are 4, 8, 12, 16 and 24 numbers of neighbors for 1, 1,
2013). In India, 5129.4 million ton potatoes (Anonymous, 1.5, 2.0 and 3.0 radii respectively. Alhindi et al., (2018)
2018) are produced for the year 2017-18, and it occupied stated that output feature has a dimension of 1182 by using
third position in the world (Anonymous, 2017). In India, radius and number of neighbors as 4 and 14 respectively. In
potato production has increased substantially over the last histogram of oriented gradients (HOG), an image is divided
six decades. The analysis shows that the potato yield and into number of cells by using different cell sizes and draw
production increased respectively at the national level at gradients over the image. Features of HOG and LBP are
1.10 and 5.98 million per year during the last decade (Bai et used in facial recognition. An output of 1224 bin dimension
al., 2009). In this paper, gray level co-occurrence matrix HOG vector was formed by using cell size of 18-by-18 and
(GLCM), local binary patterns (LBP), Histogram of block size of 1-by-1. Gabor features are used for detecting
Aloriented gradients (HOG), gabor and radon feature cancerous images from non cancerous images (Alhindi et
extraction techniques were appiled to crack, rotten, sprout, al., 2018). Radon features uses projections to calculate
skin peel and good potato images in adapted and improved feature vector (Alhindi et al., 2018).
techniques, because it is very necessary and important to
find best technique interms of feature size and time taken to
III. MATERIALS AND METHODS
extract features.
This section starts with image capturing system
development. Five adapted, improved feature extraction
II. REVIEW OF LITERATURE techniques used in image analysis were compared and
Local binary patterns and gray level co-occurrence conclusions were drawn.
matrix properties are used to estimate the textural features 3.1 Setup for Image Capturing
(Matlab, R2018b). Moallem et al., (2013) used three
The image processing units were consisting of
properties of gray level co-occurrence matrix such as
acquisition devices, followed by image analysis unit like
contrast; correlation and energy are used for classification
computer with image storage unit as shown in Figure 1.
using support vector machine and artificial neural network.
Image capturing requires an artificial light, image sensor
Alhindi et al., (2018) given that local binary patterns (LBP)
and image storage unit as shown in Figure 1.
for recognizing mammography images, facial recognition
and for extracting features of cancerous areas. Ojala et al.,

Fig.1 Image processing setup developed for potatoes

A setup was developed with the help of a with inbuilt six LED lights and potentiometers were
40×40×40 cm box using a 20-gauge galvanized iron sheet. installed on the inner four side walls of the box focussing in
Four still web cameras (Model no. QHM495LM) equipped the centre of the box (Figure 1). The four ends of cameras

ISSN: 2456-1878 (Int. J. Environ. Agric. Biotech.)

https://dx.doi.org/10.22161/ijeab.94.22 162
Babu et al. Traditional Low-Efficient and Anti-Ecological Agriculture Must Be Replaced with Modern Cellular and
Hydroponic Agriculture

were connected to laptop through USB (Zebronics, ZEB- (Ojala et al., 2002). For calculating the plot time between
04HB) Hub. While capturing care must be taken by squared error and histogram bins, images are rotated to 30°.
allowing light to settle on potatoes. Potentiometers were For calculating feature vector rotation invariant property
used for adjusting the light intensity. Walls of the box were (Upright) is set to ‘true’ and incase of calculating plot, time
painted with black color to eliminate reflections. A door was rotation invariant property is set to ‘false’. The following
provided at the front to place and remove potatoes. After algorithm was developed for local binary patterns.
developing setup, next step is image capturing. Crack, Algorithm 2: Extracting local binary pattern (LBP) features
rotten, sprout, skin peel and good potato images each of
Input: Color potato image
twenty having different image sizes are taken for adapted
and improved feature extraction. Feature extraction is the Output: Local binary feature vector and plot between
step before classification. squared error and LBP histogram bins
3.2 ADAPTED FEATURE EXTRACTION Step 1: Read an image to a variable using imread function
TECHNIQUES Step 2: Convert the color image into gray image using
Feature extraction techniques like gray level co- rgb2gray function
occurrence matrix properties, local binary patterns, Step 3: Rotate gray image to an angle of 30°
histogram of oriented gradients, gabor and radon transforms
Step 4: Calculate local binary feature vector of gray image
were applied to five types of potato images by some default
and rotate gray image using extractLBPFeatures
values and adapting changes according to their properties
function
(Matlab, R2018b). Features from images are extracted to
maximise the object identification and information Step 5: Square the difference of local binary feature vectors
available in images. To calculate time taken to plot for all of gray image and rotate gray image
techniques except GLCMP, tic function at starting of the Step 6: Draw bar chart between squared error and LBP
algorithm and toc function at the end of the algorithm is histogram bins
applied (Matlab, R2018b).
3.2.3 Adapted Histogram of Oriented Gradients
3.2.1 Adapted Gray Level Co-occurrence Matrix
In adaptive method all default values are used to
Properties (GLCMP)
calculate HOG feature vector i.e. Cell size 8-by-8, block
In adaptive method, an offset of [0 1] is applied to size 2-by-2, block over lap 1-by-1, number of bins 9 and
find the contrast, correlation, energy and homogeneity Used signed orientation is set to false (Dalal and Triggs,
values (Haralick et al., 1973). Following algorithm was 2005). Algorithm for determining of histogram of oriented
developed for extracting the properties of GLCM. gradients is given below:
Algorithm 1 : Finding properties of GLCM Algorithm 3: Feature vector of histogram of oriented
Input: Color potato picture gradients
Output: Values of gray level co-occurrence matrix i.e. Input: Gray Image
variance, correlation, uniformity and Output: Visualization of Gradients and feature vector
homogeneity
Step 1: Read an image using imread function
Step 1: Read color potato image to a variable using imread Step 2: Selecting the size of cell
function. Step 3: Extract HOG Feature vector using
Step 2: Convert the color potato image to gray image. extractHOGFeatures function.
Step 4: Plot distribution of gradients
Step 3: Specify number of gray levels and offsets to observe
Step 5: Displaying result.
GLCM size using graycomatrix.
3.2.4 Adapted Gabor Feature
Step 4: Find GLCM properties using graycoprops at given
number of offsets. A wavelength of 2 (Matlab, R2018b) to 10 with
equal interval of 2 and orientation of 0 to 180 with an
Step 5: Find the mean of above set of values to calculate
interval of 22.5 degrees was taken for extracting adapted
single value of variance, correlation, uniformity
gabor feature vector. Five wavelength values and eight
and homogeneity.
orientation values are suggested by (Bai et al., 2009 and Zhu
3.2.2 Adapted Local Binary Patterns et al., 2007). Algorithm for extracting gabor feature is given
In adapted method, the default values of 1 and 8 below:
are taken as the radius and number of neighbors respectively Algorithm 4: Extracting gabor features

ISSN: 2456-1878 (Int. J. Environ. Agric. Biotech.)

https://dx.doi.org/10.22161/ijeab.94.22 163
Babu et al. Traditional Low-Efficient and Anti-Ecological Agriculture Must Be Replaced with Modern Cellular and
Hydroponic Agriculture

Input image: Color image respectively. Parameters like improved radon are selected
Output image: Gabor filtered image and visualization of based on preliminary studies. Except for GLCMP, time
wavelength and orientation taken to execute plot was applied.

Step1: Read the color image using imread function 3.3.1 Improved Gray Level Co-occurrence Matrix
Properties (GLCMP)
Step2: Resize the image by using a fraction of 0.1 using
imresize function In improved gray level co-occurrence matrix size
an offset of [4 0] was applied to find the feature vector
Step3: Convert resized color image into gray scale image
containing contrast, correlation, energy and homogeneity
using rgb2gray function
values. Offset parameter [4 0] taken because of good
Step4: Assign the wavelength and orientation values coverage of neighboring pixels in calculating contrast,
Step5: Form gabor array correlation, energy and homogeneity.

Step6: Apply gabor filter to gray scale image using 3.3.2 Improved Local Binary Patterns (LBP)
imgaborfilt function In improved method of LBP the radius and number
Step7: Display gabor filtered images of neighbors taken as 4 and 14 respectively (Alhindi et al.,
2018) and the remaining steps are same as shown in
Step 8: Give wavelength and orientation ranges to visualize
Algorithm 2.
the wavelength and orientation of spatial kernels.
3.3.3 Improved Histogram of Oriented Gradients
3.2.5 Adapted Radon Feature
(HOG)
In adaptive radon transform the default values of
In improved histogram of oriented gradients a cell
theta ranges from 0 to 180° are taken (Matlab, R2018b).
size of 18-by-18 and block size of 2-by-2 was used (Alhindi
Algorithm for extracting radon features was given below:
et al., 2018) instead of 8 by 8 in adapted HOG, while
remaining steps were same.
Algorithm 5 : Calculating radon features of a color
3.3.4 Improved Gabor Features
potato image
In improved gabor feature extraction four
Input: Potato color (RGB) image
wavelength values and four orientation values are taken.
Output: Radon features of an image Wavelengths of 32, 26, 18 and 16 pixels/cycle and
Step 1: Read a color image to a variable using imread orientation values of 0, 45, 90 and 135 degrees are taken
function (Fogel and Sage, 1989) and remaining values is similar to
Algorithm 4.
Step 2: Converting RGB image into gray scale image using
rgb2gray function 3.3.5 Improved Radon Features
Step 3: Set projection angle theta (θ) between 0 to 179°. In improved radon feature extraction the radon
transforms are calculated by using theta value ranging from
Step 4: Find the sizes of radon and radial coordinates
0 to 90° and remaining values similar to Algorithm 5.
feature vectors
Step 5: Plotting graph between projection angle (theta) and
vector coordinates of radon using plot function IV. RESULTS AND DISCUSSION
The developed image capturing setup has a
3.3 IMPROVED FEATURE EXTRACTION
resolution of 0.22 mm/pixel which was in acceptance with
TECHNIQUES
strawberry grading system with 0.17 mm/pixel resolution
Adapted feature extraction techniques were (Xu and Zhao, 2010). The results of five feature extraction
limited to their standard type of algorithm but improved methods in terms of adapted and improved techniques were
feature extraction techniques are will give better and compared and analysed. The size of GLCM in adapted and
convenient results than adapted techniques. Contrast, improved techniques is same i.e. 8-by-8, because all input
correlation, energy and homogeneity values were computed images were 8-bit images.
by changing offset in improved GLCMP (Gray Level Co-
4.1 Comparison of adapted and improved gray level co-
occurrence Matrix Properties) technique. Cell size and
occurrence matrix properties
block size are selected to find improved HOG technique.
Wavelength and orientation values are selected to find Similar patterns are observed for properties of
improved gabor features. Half of the projection angle used GLCM among all potato classes for adapted and improved
in adapted technique used to find radon features feature extraction techniques. Contrast values are lower in

ISSN: 2456-1878 (Int. J. Environ. Agric. Biotech.)

https://dx.doi.org/10.22161/ijeab.94.22 164
Babu et al. Traditional Low-Efficient and Anti-Ecological Agriculture Must Be Replaced with Modern Cellular and
Hydroponic Agriculture

adapted method for all potato classes compared to improved it is 0.04, for sprout potatoes it is 0.035, skin peel potatoes
method. A difference of contrast values in adapted and it is 0.04 and for good potatoes it is 0.025. Energy values
improved method was observed in the range of 0.2-0.3. are 0.2 times higher in adapted method for all potato classes
Correlation values are higher in adapted method for all compared to improved method. Homogeneity values are
potato classes compared to improved method. A difference higher in adapted method for all potato classes compared to
of correlation values between adapted and improved improved method (Figure 2).
method of crack potatoes is 0.04, while for rotten potatoes

Fig.2 Comparison of adapted and improved GLCMP in relation to potato classes

Homogeneity value of 0.06 was observed for crack potato depend upon the number of elements in the wavelength and
as difference value of adaptive and improved methods. orientation vectors. The size of gabor feature array is the
Similarly, a difference of 0.069 was observed for rotten, product of number of elements in the wavelength and
0.07 for sprout, 0.0728 for skin peel and 0.063 for good orientation vectors. Gabor features in adapted and improved
potatoes were observed. techniques were compared using feature vector lengths and
4.2 Comparison of Adapted and Improved Gabor time taken to plot magnitude and spatial kernels in seconds.
Features An improved gabor method has lower gabor feature length
compared to adapted gabor method (Figure 3).
Adapted gabor feature vector length (40) was same for all
types of potato images. The size of gabor feature length

Fig.3 Comparison of adapted and improved gabor feature vector lengths in relation to potato classes

ISSN: 2456-1878 (Int. J. Environ. Agric. Biotech.)

https://dx.doi.org/10.22161/ijeab.94.22 165
Babu et al. Traditional Low-Efficient and Anti-Ecological Agriculture Must Be Replaced with Modern Cellular and
Hydroponic Agriculture

Fig.4 Comparison of adapted and improved plot timings of gabor magnitude and spatial kernels in relation to potato classes

Improved gabor method requires less time to plot gabor 4.3 Comparison of Adapted and Improved HOG
magnitude and spatial kernels for all potato classes (Figure Features
4). The HOG feature vector lengths in adapted HOG
The time taken to plot magnitude and real spatial parts of all method had highest compared to the improved HOG
types of potato images in adapted radon method is twice to method proposed by (Alhindi et al., 2018) (Figure 5).
the time taken in improved radon method.

Fig.5 Comparison of HOG feature vector lengths in adapted and improved techniques in relation to potato classes

For crack and rotten images, an improved HOG method has more HOG visualization time because of cell
method proposed by Alhindi et al., (2018) has required size 8- by-8 compared to improved HOG cell size of 18-by-
higher time to plot visualization than adapted HOG method. 18 (Figure 6).
While for sprout, good and skin peel images, adapted HOG

ISSN: 2456-1878 (Int. J. Environ. Agric. Biotech.)

https://dx.doi.org/10.22161/ijeab.94.22 166
Babu et al. Traditional Low-Efficient and Anti-Ecological Agriculture Must Be Replaced with Modern Cellular and
Hydroponic Agriculture

Fig.6 Comparison of adapted and improved HOG visualization times in relation to potato classes

Fig.7 Comparison of adapted and improved techniques in relation to potato classes using LBP feature vector lengths

4.4 Comparison of Adapted and Improved LBP The mean time to plot squared errors in adapted and
Features improved methods for crack images were found to be
The feature vector lengths of an improved method 0.6378 s and 0.6305 s respectively, for rotten images 0.2098
proposed by Alhindi et al., (2018) were higher (185) s and 0.2622 s, for sprout images 0.1911 s and 0.2209 s, for
compared to adapted method (59) (Figure 7). skin peel images 0.2197 s in both methods, for good potato
images 0.2672 s and 0.2565 s (Figure 8).

ISSN: 2456-1878 (Int. J. Environ. Agric. Biotech.)

https://dx.doi.org/10.22161/ijeab.94.22 167
Babu et al. Traditional Low-Efficient and Anti-Ecological Agriculture Must Be Replaced with Modern Cellular and
Hydroponic Agriculture

Fig.8 Comparison of adapted and improved techniques in relation to potato classes using squarred errors in LBP

4.5 Comparison of Adapted and Improved Radon transform plotting time. From radon feature sizes of adapted
Features and improved methods, it is inferred that the theta value
Radon feature sizes of adapted (Matlab, R2018b) effect the size of the radon feature (Figure 9).
and improved methods were same but differ in radon

Fig.9 Comparison of radon feature vector lengths of adapted and improved techniques in relation to potato classes

At half of the theta value (90°) in improved compared to adapted method (at theta value 180°) (Figure
method, the time taken to plot radon transforms is lower 10).

ISSN: 2456-1878 (Int. J. Environ. Agric. Biotech.)

https://dx.doi.org/10.22161/ijeab.94.22 168
Babu et al. Traditional Low-Efficient and Anti-Ecological Agriculture Must Be Replaced with Modern Cellular and
Hydroponic Agriculture

Fig.10 Comparison of time taken to plot radon transforms of adapted and improved techniques in relation to potato classes

V. CONCLUSIONS REFERENCES
The adapted feature vector lengths in HOG were [1] Alhindi, T.J., Kalra, S., Ng, K.H., Afrin, A, and Tizhoosh,
found highest compared to the improved HOG method for H.R. 2018. Comparing lbp, hog and deep features for
all potato classes. For crack and rotten potato images, HOG classification of histopathology images. In: Proceedings of
the IEEE World Congress on Computational Intelligence
visualization time in improved method was higher
held at Rio de Janeiro in 3-8 July, 1-7.
compared to adapted method but for remaining potato
[2] Anonymous. 2018. Agricultural and processed food products
image classes adapted method has more HOG visualization export development authority.
time compared to improved method. The LBP feature http://apeda.in/agriexchange/India%20Production/India_Pro
vector lengths in improved method were higher compared ductions.aspx?cat=Vegetables&hscode=1083.
to adapted method for five potato classes. In improved LBP [3] Anonymous. 2017. https://www.potatopro.com/india/potato-
method, the squared error’s for all potato classes were same. statistics.
The gabor feature vector length was highest in adapted [4] Bai, Y., Guo, L., Jin, L. and Huang, Q. 2009. A novel feature
method compared to improved method for all potato extraction method using pyramid histogram of oriented
gradients for smile recognition. In: Proceedings of IEEE
classes. An Improved Gabor method required less time to
International Conference on Image Processing held at Cairo
plot gabor magnitude and spatial kernels for all potato
in 7-10 November, 3305-3308.
classes compared to adapted method. For all potato classes [5] Dalal, N. and Triggs, B. 2005. Histogram of oriented
radon feature vector sizes were same in both adapted and gradients for human detection. IEEE Computer Society
improved methods but differ in plotting times of radon Conference on Computer Vision and Pattern Recognition
transformation. Theta value affected the radon feature [6] Fogel, I. and Sagi, D. 1989. Gabor filters as texture
vector size in all potato classes. In improved radon method, discriminator. Biological Cybernetics 61: 103-113.
less time had taken for plotting radon transforms for five [7] Haralick, R.M., Shanmugam, K. and Dinstein, I. 1973.
classes of potato compared to adapted method. In improved Texural features for image classification. IEEE Transactions
on Systems, Man, and Cybermetics. 3: 610-621.
GLCMP method, contrast values were highest compared to
[8] MATLAB, R2018b. Help documentation. Ver. 9.5.0.944444.
adapted method for all five potato classes. In adapted
Licence Number: 968398. The Mathworks, Inc. USA.
GLCMP method, correlation values were greater than [9] Moallem, P., Razmjooy, N. and Ashourian, M. 2013.
adapted method for all potato classes. In adapted GLCMP Computer vision based potato defect detection using neural
method, energy values were greater for all potato classes networks and support vector machine. International Journal
compared to improved method. Also, in adapted GLCMP of Robotics and Automation 28: 1-9.
method, homogeneity values were greater for all five potato [10] Ojala, T., Pietikainen, M. and Maenpaa, T. 2002. Multi-
classes compared to improved GLCMP method. resolution gray-scale and rotation invariant texture
classification with local binary patterns. IEEE Transactions
on Pattern Analysis and Machine Intelligence 24: 971-987.

ISSN: 2456-1878 (Int. J. Environ. Agric. Biotech.)

https://dx.doi.org/10.22161/ijeab.94.22 169
Babu et al. Traditional Low-Efficient and Anti-Ecological Agriculture Must Be Replaced with Modern Cellular and
Hydroponic Agriculture

[11] Rani, S. and Prasoon, M. 2013. Analysis of potato production

performance and yield variability in India. Potato Journal 40:
38-44.
[12] Xu, L. and Zhao. 2010. Automated strawberry grading
system based on image processing. Computers and
Electronics in Agriculture 71S : S32-S39.
[13] Zhu, B., Jiang, L., Luo, Y. and Tao, Y. 2007. Gabor feature-
based apple quality inspection using kernel principle
component analysis. Journal of Food Engineering 81: 741-
749.

ISSN: 2456-1878 (Int. J. Environ. Agric. Biotech.)

https://dx.doi.org/10.22161/ijeab.94.22 170