2016 IEEE-fruits-mango-Methods in Defect Classification of Mangoes Using-Ashok2016

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

2016 Second International Conference on Cognitive Computing and Information Processing (CCIP)

A Comparative Study of Feature Extraction


Methods in Defect Classification of Mangoes using
Neural Network

VaniAshok D.S.Vinod
Computer Science & Engineering Information Science & Engineering
Sri Jayachamarajendra College of Engineering Sri Jayachamarajendra College of Engineering
Mysore, India Mysore, India
vanisj@sjce.ac.in dsvinod@daad-alumni.de

Abstract— The “king of fruits” Mango (Mangifera indica L.) is mango producing countries. India is the predominant supplier of
the most sought after fruit for both direct and indirect consumption mangoes to the west Asian market and is estimated to worth
across the globe. Since it has very high export value, there is a need more than 100 crore. Some of the prominent mango varieties
to develop a technique that is capable of classifying the defects of cultivated in different states of India are Bombai, Bombay Green,
mangoes objectively. Any classifier performance is dependent on Himsagar, Kesar, Suvernarekha, Alphonso, Mankurad,
the features extracted from the region of interest of the sample. In Bangalora, Vanraj, Banganapalli, Dashehari, Langra, Kishen
this paper, a comparative study of feature extraction methods is Bhog, Zardalu, Mankurad, Fazli, Fernandin, Mulgoa, Neelum
made to classify the visible defects of Mangoes. “Alphonso” mango and Chausa.
cultivar was chosen for the experimentation. 1766 color images
with different quality classes were acquired, pre-processed and
textural features were extracted considering one feature at a time A. Defects Occuring in Mangoes
and also in combination for each color image. Hence, we obtained 9 The quality indices of mangoes are uniformity of shape and
different cases of different textural features combination. size, absence of decay and defects, skin color which is specific to
Furthermore, most relevant features were selected from each case a particular variety, flesh color, firmness (juiciness, fiber content)
using sequential forward selection algorithm. The textural features and flavor (sweet, acidity, aroma intensity) [3]. Mangoes are
like statistical, LBP and filter banks were found to be effective in susceptible to physical and physiological defects. The physical
designing neural network (NN) using generalized linear model defects that appear during pre-harvesting, harvesting and post-
classifier with cross validated performance accuracy of 90.09%, harvesting stages are as follows: Anthracnose, jelly seed, insect
90.26% and 90.26% for linear, logistic and softmax activation
damage, lenticel damage, scab, scars, skin breaks and cracks, soft
functions respectively.
nose, stem-end cavity, sunburn, sunscald, bruising, decay,
Keywords—Defects in Mango; Image segmentation; Texture immature, overripe, shriveling.
Features; Neural Network ; Validation Postharvest practices and treatments which are normally
carried out to increase the shelf life, retain the quality, nutrition
I. INTRODUCTION and functional compositions of mangoes also induce significant
defects and damages to the fruits. Some of the post-harvest
The quality monitoring of agricultural and food products has
practices are: 1) Low temperature, where the storage life of
taken a new dimension due to the increasing quality awareness
certain mango cultivars is extended to 2 to 3 weeks by storing at
among consumers. Quality assessment of fruits and vegetables is
10o-15oC temperature range, 2) Ionization radiation, where the
done based on the analysis of external features like color, size,
mangoes are irradiated with the application of gamma rays,
shape, texture and defects/blemishes [1]. Although consumers are
electrons, X-rays, microwaves during storage, to minimize
influenced to select fruits based on their appearance, other
decaying and 3) Usage of chemicals, where mango ripening can
parameters like firmness, aroma and presence or absence of
be retarded upto 6 days by spraying Gibberellic acid prior to
defects play an important role in determining the quality of fruits.
harvest thereby increasing the shelf life [4].
The “king of fruits” Mango (Mangifera indica L.) has very high
nutritive value and rich in carotenes. There are hundreds of These post-harvest practices and storage methods have to be
mango cultivars distributed throughout the world, of which Asia carried out at correct specifications. Otherwise they induce
and India have over 1000 classified varieties [2]. India produces physical defects in fruits such as chilling injury, sunken shoulder
50% of total world’s mango and ranks first among world’s areas, external discoloration, uneven ripening etc. These defects

978-1-5090-1025-7/16/$31.00 ©2016 IEEE


are prominently visible on the affected fruits and have to be and a linear regression analysis was used to analyse the
segregated from the healthy ones before marketing. Traditionally relationship between 2-dimensional area of thresholded image
this is done by laborers who inspect the fruits manually. This is and mango weight. Nagle, Marcus, et al. [10] have suggested a
time consuming and subjective in nature. The process of sorting non-destructive Thai mango quality evaluation based on the color
tons of mangoes manually is slow, costly and inaccurate process. of fruits. Physical defects like anthracnose and latex stain were
Hence there is scope for developing computer vision based induced on the fruits artificially. The image acquisition was done
sorting system for automatic inspection of defects in mangoes. in RGB color space and then transformed into L*a*b* and the
defects were identified using different segmentation techniques.
The physiological disorders are related to various internal A. M. Aibinu et al. [11] proposed a technique to sort the fruits
flesh breakdown such as spongy tissue (commonly found in such as apple, banana and mango. The acquired images were
“Alphonso” variety), soft nose, jelly seed, insidious fruit rot etc. segmented, then the shape information was extracted using
In all these cases there are no external symptoms on the affected Fourier descriptors and features were estimated by spatial domain
fruit and the defects are noticeable only when the fruit is cut analysis. The artificial neural network was used to detect the fruit
open. Hence some non-destructive methods have to be devised to based on color.
separate the affected fruit from the lot during marketing and
exporting. The mangoes infested with pests are also of great Depending on the statistical textural features, Jagadeesh
concern as they incur losses to growers and exporters. The Devdas Pujari et al. [12] have proposed an algorithm to grade and
Mango seed weevil and fruit flies are major quarantine pests due classify anthracnose fungal disease in mango, pomegranate and
to which there is possibility of banning mango fruit imports in grape fruits. Various segmentation techniques were employed to
some countries. isolate the anthracnose affected areas from the normal area and
affected areas were graded by computing the fraction of affected
A number of on-line automatic defects/diseases recognition area. In the next phase, textural features were extracted and
systems have been proposed for various fruits in recent years. A BPNN classifier was used with an accuracy of 84.65% and
non-invasive inspection method has been reported by Leiva 76.6% for normal type and anthracnose affected type
Valenzuela [5] to assess the internal quality attributes and respectively.
external features of blueberries using hyperspectral imaging
technique. The blueberries were classified into two classes of
firmness and soluble solid contents (SSC) of the fruits were B. Feature Extraction Techniques
predicted. Similarly, a non-destructive technique like soft X-ray Feature extraction is the method of quantifying the significant
was used to determine the internal defects of pomegranate and features of the image. It is the crucial step in image classification
classification algorithm was proposed by Payel G and C. K. Sunil where image properties are analysed and quantified features are
to classify defective pomegranate from the non-defective ones organized into classes i.e., an image is classified according to its
[6]. The X-ray images were converted to grayscale and then to contents [13]. Features extracted from each image are divided
binary image using adaptive thresholding to remove background into geometrical features and intensity features.
and then Blob analysis block was done to calculate centroid, label Geometrical features include location and size features such
matrix and blob count. The area calculation algorithm was also as area, perimeter, orientation, Euler number, center of gravity,
proposed to determine the value of the exact defected area of solidity, shape information invariant of a region and its
every fruit. neighbourhood using Fourier descriptors and invariant moments
A computer vision methodology to assess the internal quality [14].
classification of sliced chestnuts was proposed by Irwin R. Donis Intensity/Color features provide information about the
Gonzalez et al. [7]. 1790 color images from slices with different intensity of the region in gray and different color spaces like R,
quality classes were acquired, preprocessed and color, G, B, L*, a*, b* etc and its neighborhood. They include standard
geometrical and texture features were extracted. The relevant intensity features such as mean, standard deviation, kurtosis,
features were selected using a sequential forward selection mean gradient, mean Laplacian, texture information based on
algorithm and quadratic discriminant classifier was used to Haralick approach, Local binary patterns (LBP), and contrast
classify the chestnut slices into 5 classes. An automatic algorithm features [15].
was developed by Vahid Mohammadi et al. [8] to grade the
persimmon fruits into three maturity stages based on the skin In this paper we present a comparative study of the feature
color. After extracting the color features, linear and quadratic extraction methods for defect classification in mangoes. The
discriminant analysis classifiers were used to determine the performance of the data models obtained from different
suitability of the vision system with an accuracy rate of 90.24%. combination of feature extraction methods are evaluated with
Neural Network using a Generalized Linear Model (NNGLM)
Only a few works have been reported on the automatic considering various output activation functions. The
quality evaluation of mangoes when compared to other fruits classification accuracy rate is calculated to measure the
such as apple, pomegranate, watermelon etc. Weight estimation performance of the system. The experimentations were
of Chokanan mangoes depending on the size of the fruit using conducted on “Alphonso” mangoes as it is the most sought after
image processing techniques has been reported by C.C. Teoh et mango cultivar for direct consumption.
al. [9]. A total of 100 mango images were acquired, preprocessed
II. MATERIALS AND METHODS

The proposed model for automatic defect classification of


mangoes consists of the following steps: (1) Sample collection
and preparation, (2) Image acquisition and pre-processing, (3)
Segmentation, (4) Feature extraction and selection, (5)
Classification and validation.

A. Sample Collection and Preparation


“Alphonso” mangoes were acquired from an orchard near
Mysuru, Karnataka from the 2015 and 2016 harvests. The fruits
were visually inspected and based on appearance and surface Fig. 1. Overall sample distribution used to train the classifier.
defects the fruits were divided into two classes. Healthy mangoes
were defined as normal sized, well-colored, free from
irregularities such as defects, spots, compressions, wrinkles and
impacts. Fruits with various physical defects such as lenticel
damage, scars, skin breaks, scab, cracks, stem-end cavity,
sunburn, sunscald, soft nose, bruising, decay, immature, overripe,
shriveling, sapburn were defined as defective mangoes. Finally, (a)
to create a third class, chilling injured mangoes, a batch of
healthy, defect-free mangoes were selected and chilling injury
was induced by storing the fruits at 4o-5oC for about 25 days.
Chilling injury is a visible defect that depends on time and
temperature. The symptoms consist of greyish scald like
(b)
discoloration of skin, uneven ripening, skin pitting and reduction
in the levels of aroma and flavor during ripening [4].

B. Image Acquisition and Pre-processing


Images were acquired using Sony Cyber-shot DSC-WX7 (c)
digital camera. A total of 1766 color images were acquired with a Fig. 2. Sample fruit images for each of the 3 categorical classes. (a) Chilling
resolution of 640 X 480 pixels and 60.5KB size. The image injured mangoes (b) Defective mangoes (c) Healthy mangoes
acquisition was done between 12PM to 3PM under natural light. When compared to other color spaces, YCbCr facilitates more
Camera was positioned so as to avoid the formation of fruit effective segmentation by clearly differentiating fruit from the
shadow and a white background was used to ease the background [16]. Hence the YCbCr color space was chosen to
segmentation process. The images were stored in JPG format. compute global threshold segmentation. It was performed in two
Inorder to reduce the computational complexity, in the pre- steps. The first step consisted of building a binary mask by
computing thresholding on Cb channel information of the image
processing stage, the acquired images were cropped and resized
and the second step is to perform morphological operations on
to a dimension of 200 X 150 pixels and 5.58 KB size. After
the segmented image to remove isolated pixels in the background
further visual inspection, the samples were labeled and the class and to fill holes in the foreground. An example for this
of each sample was determined. The quality of fruits is segmentation is shown in Fig. 3
categorized into one of three classes: “healthy”, “defective” and
“chilling injured”. Among 1766 images, 742 images were labeled 1000

as chilling injured mangoes (class 1), 552 images were labeled as 800

defective mangoes (class 2) and 472 images were included under


600

400

healthy mangoes (class 3). The overall sample distribution that is 200

used to train the classifier is shown in Fig. 1 and sample fruit


0

0 50 100 150 200

images in three classes are shown in Fig. 2. (a) (b) (c) (d)
Fig. 3. Segmentation of fruit image. (a) Original image (b) Global thresholding
C. Segmentation obtained from histogram (c) Histogram (d) Segmented image
Image segmentation was performed to analyze the
information necessary for fruit defect assessment i.e., it was D. Feature Extraction/Selection
carried out to isolate the foreground (fruit) from the background After the image segmentation, feature extraction was
(the rest of the image). performed to obtain the quantitative attributes of each sample
from the regions of interest. Color components were extracted depends on the number of classes. The generalized linear model
from each color image in red, green and blue (RGB) channels was trained using the Iterative Reweighted Least Squares (IRLS)
and in addition, gray scale image was obtained from each image. algorithm to set the weights. This was found to be a more
Therefore, four intensity images were obtained from each color efficient alternative to a non-linear optimization routine [17]. 10-
image. The extracted features included standard intensity fold cross-validation was used to evaluate the performance of
features, local binary patterns (LBP), Discrete Fourier the classifier with a confidence interval of each classification
Transforms (DFT) and Discrete Cosine Transform (DCT), Hu being 95%. In the cross validation, 90% of the samples were
with intensity and Gabor filters. The features that depend on used for training and 10% were used for 10 validation
contrast, orientation and location were avoided as the fruit replications.
quality is independent of these characteristics. Extracting the
features upon which the fruit quality does not depend on leads to III. RESULTS AND DISCUSSIONS
false correlation. Each mango image in the dataset is decomposed into R, G, B
The extracted standard chromatic features for RGB color color channels and the grayscale for extracting textual features.
space and grayscale include; Standard intensity information for In this study, SFS with the Fisher discriminant objective function
four intensity channels related to mean, kurtosis, standard was used to reduce the extracted large features. Computational
deviation, skewness, mean gradient (first-order derivative) and time decreases by reducing the overall feature space and this
mean Laplacian (second-order derivative); Linear binary allows for possible use in real-time implementation. Instead of
patterns which are influential features for texture classification considering all the intensity features together, a comparative
were extracted to compute the texture through the relationship study has been done by considering intensity features one at a
between the intensity of each pixel and eight neighboring pixels, time. The feature extraction and selection followed in this study
can be divided into several cases and classifier performance on
using the occurrence histogram [17]; Intensity Gabor textural
each of these cases was analysed. The intensity features in each
features, extracted from the magnitude filtered images with
case, total number of extracted features and the computational
different scales and orientations are helpful for extracting useful time of feature extraction process is given in Table I.
textural features from an image [7]; Filter banks extracted
texture information from image transformations like DFT and
magnitude, phase through DCT; intensity information based on TABLE I. COMBINATION OF INTENSITY FEATURES FOR DEFECT
CLASSIFICATION IN MANGOES
Hu moments. The features were extracted using “Balu” toolbox
for recognition patterns. These features were extracted Extracted Computational
Case Intensity Features used
Features Time in Minutes
separately and also in some combinations to explore which of Standard intensity
the features combinations increases the performance of the features and textural
1 216 24.18
classifier the most. features from Fourier and
After feature extraction, feature selection was done using DCT
Standard intensity
“Sequential Forward Selection” strategy with objective function 2
features and LBP
260 23.56
of Fisher Discriminant (SFS-Fisher). The SFS-Fisher selects a Standard intensity
small subset of features from total features that leads to the 3 features and Hu with 52 23.29
minimal classification error. In the space of selected features, it intensity filters;
guarantees minimum intraclass and maximum interclass Standard intensity
4 features and Gabor 292 79.0
variations [14]. filters;
Standard intensity
E. Classification and Validation 5 features, LBP and Fourier 452 28.0
With optimized selected features, the classification of and DCT;
mangoes into different classes was carried out by Neural Standard intensity
6 features, LBP and Gabor 528 75.34
Network (NN) using generalized linear model. Since the NN filters;
classifier was found to achieve higher performance for Standard intensity
multiclass classification problems with imbalanced training data, 7 features, LBP and Hu 288 23.29
NN architecture was considered for the proposed work [18]. The with intensity;
Standard intensity
performance of the classifier was evaluated considering different features, LBP, Fourier
output activation functions like linear, logistic and softmax. A 8 480 26.42
and DCT and Hu with
two-layer NN with linear activation function, a three-layer NN intensity;
using logistic activation function and a three-layer NN using Standard intensity
features, LBP, Fourier
softmax activation function were used to classify the objects (in 9 and DCT, Hu with 748 80.4
this case, four images per sample) to a specific class. The intensity and Gabor
number of neurons in the input layer corresponds to the number filters.
of input features and the number of neurons at the output layer
TABLE II. CLASSIFICATION ACCURACY RATE WITH ELAPSED TIME IN SECONDS FOR 9 FEATURE EXTRACTION CASES
Classifier Case 1 2 3 4 5 6 7 8 9

Selected features 16 19 19 15 17 19 19 20 20
Neural Network
Accuray 89.42 89.53 88.73 89.71 89.76 89.58 89.18 89.98 90.09
(Linear)
Time 1 1 1 1 1 1 1 1 1

Selected features 19 18 20 20 20 20 19 20 18
Neural Network
Accuray 90.55 89.74 89.53 90.14 90.04 90.25 89.63 90.60 90.26
(Logistics)
Time 4 3 4 4 4 4 5 5 4

Selected features 17 17 13 20 20 17 20 20 20
Neural Network
Accuray 90.33 89.81 89.45 89.91 89.80 90.20 90.13 90.67 90.26
(Softmax)
Time 137 139 135 137 136 138 140 138 143

TABLE III. COMMON BEST FEATURES FOR CASE 1 AND CASE 9


time for the classification are as shown in Table II. From Table I
No. of we can infer that the total number of features as well as the
Case 1 Case 9
Features
g-Mean Boundary g-Mean Boundary
computational time increases with the number of intensity
1
Gradient Gradient features. But since intensive feature exploration process is done
B-Mean Boundary g-Gabor-J offline, it will not influence the elapsed time of feature selection
2
Gradient and classification process. The achieved performances are
3
R-Intensity StdDev B-Mean Boundary summarized in Table II. The performance accuracy rate for
Gradient various intensity features is maximum for all three classifiers
B-Intensity Skewness R-Intensity StdDev
4 (90.09%, 90.26% and 90.26%) for Case 9, which includes all five
R-Mean Boundary B-Intensity Skewness intensity features, than the other cases having less number of
5
Gradient intensity features. But we also observed that, the accuracy of the
g-Intensity StdDev R-Mean Boundary classifier for Case 1(containing Standard intensity features and
6
Gradient
textural features from Fourier and DCT) is more than the
7 G-Intensity StdDev g-Intensity StdDev
accuracy rate of Case 9 where all five intensity features (Standard
8 G-Intensity Mean G-Intensity StdDev intensity features, LBP, Fourier and DCT, Hu with intensity and
G-Intensity Skewness B-Hu-moment-int 4
Gabor filters). Hence, instead of extracting all the intensity
9 features and then performing classification, the same
10 B-Intensity Kurtosis B-Gabor-J performance can be achieved by considering only a few,
G-DCT(1,3) g-LBP(1,27)[8,u2]
significant texture features.
11
g-DCT(1,3) B-LBP(1,56)[8,u2] Table III indicates the common best features selected for both
12 case 1 and case 9 for all three NN activation functions. The best
13 R-Fourier Abs (2,1) B-Intensity Kurtosis features selected were standard intensity features and local binary
G-DCT(1,1) R-LBP(1,34)[8,u2] patterns. Standard intensity features sums up the overall
14 appearances of mangoes in the R, G, B and gray channels. LBP is
15 B-DCT(1,2) g-DCT(3,2) a power texture analysis tool which uses both statistical and
B-DCT(4,1) g-Gabor(3,6) structural characteristics of texture [19]. In our experiments we
16
found that the texture features (statistical, filter bank and LBP)
17 R-Mean Laplacian R-Intensity Skewness play a major role in the classification of mangoes based on the
g-LBP(1,36)[8,u2] defects.
18

19 G-LBP(1,59)[8,u2] IV. CONCLUSIONS


20 R-LBP(1,1)[8,u2] In this work, we have considered only the textural features for
LBP (m,n):Local Binary Patterns, where ‘m’ is the number of pixels compared defect classification as the textural features are not credited to a
with ‘n’ neighboring pixels. LBP features were rotation invariant by consideration single pixel value, but rather to set of pixels in the image and
of eight orientation values. (R – Red, G-Green, B-Blue, g-Grayscale) their relationships. In our experiments, Case 9 gives the best
The accuracy rate of the NN classifier for different output result in terms of computational time and performance for
activation functions with the best selected features and elapsed mangoes samples. And among the different output activation
functions of NN, the performance of NN with softmax activation
function achieved highest accuracy rate for almost all cases (Case
1 to Case 9) but with slightly higher computational time. NN [15] Mery, Domingo and Alvaro Soto. "Features: the more the better."
with softmax function can be chosen for the mullticlass Proceedings of the 8th conference on Signal processing, computational
geometry and artificial vision. p^ pp. World Scientific and Engineering
classification of complex data. Linear and logistic activation Academy and Society (WSEAS), Rhodes, Greece, 2008.
functions were simple and achieved good performance in [16] VaniAshok and D. S. Vinod. "Automatic quality evaluation of fruits using
reasonable time (1 to 4 seconds). Hence, it was possible to Probabilistic Neural Network approach." Contemporary Computing and
achieve reasonable classification performance with only selective Informatics (IC3I), 2014 International Conference on. IEEE, 2014.
textural features and excluding geometrical and contrast features. [17] Nabney, Ian T. "Efficient training of RBF networks for classification."
These results indicated that this study is an objective and a International Journal of Neural Systems 14.03, 201-208, 2004.
reliable tool which can be extended to in-line automatic quality [18] S. Suresh, N. Sundararajan and P. Saratchandran. "Risk sensitive loss
evaluation of mangoes. The proposed external defect functions for sparse multi-category classification problems", Information
Sciences, 178(12), 2621-2638, 2008.
classification technique combined with internal defect evaluation
using any of the available non-destructive techniques like X-ray, [19] Leiva valenzuela, Gabriel a. "Assessment of external and internal quality of
blueberries using images.", 2015.
CT, MRI etc. gives the best overall result for automatic quality
evaluation of mangoes. The classifier performance can be further
evaluated using other metrics such as precision, f-measure, recall
etc. along with accuracy rate.

REFERENCES
[1] Jha, S. N., Jaiswal, P., Narsaiah, K., Kaur, P. P., Singh, A. K., & Kumar R.
"Textural properties of mango cultivars during ripening", Journal of food
science and technology, 50(6), 1047-1057, 2013.
[2] Jha, S.N., A.R.P Kingsly, S. Chopra. "Non-destructive determination of
firmness and yellowness of mango during growth and storage using visual
spectroscopy", Biosystems Engineering 94(3), 397-402, 2006.
[3] Kader, A. A. "Mango quality attributes and grade standards: a review of
available information and identification of future research needs (report to
the National Mango Board)." Davis, CA, USA. Kader Consulting Services,
2008.
[4] Jha S. N, Narsaiah K, Sharma A. D, Singh M, Bansal S. Kumar R " Quality
parameters of mango and potential of non-destructive techniques for their
measurement – a review", J Food Sci Technol, 47(1):1–14, January–
February 2010.
[5] Leiva-Valenzuela, Gabriel A., Renfu Lu, and José Miguel Aguilera.
"Prediction of firmness and soluble solids content of blueberries using
hyperspectral reflectance imaging." Journal of Food Engineering 115.1,
91-98, (2013).
[6] Payel, G., and C. K. Sunil. "Quality Analysis of Pomegranate by Non-
Destructive Soft X-Ray Method." Journal of Food Processing &
Technology 2014, 2014.
[7] Donis-González, Irwin R., et al. "Assessment of chestnut (Castanea spp.)
slice quality using color images." Journal of Food Engineering 115.3, 407-
414, 2013.
[8] Mohammadi, Vahid, Kamran Kheiralipour, Mahdi Ghasemi-Varnamkhasti.
"Detecting maturity of persimmon fruit based on image processing
technique." Scientia Horticulturae 184, 123-128, 2015.
[9] Teoh, C. C., and AR Mohd Syaifudin. "Image processing and analysis
techniques for estimating weight of Chokanan mangoes." Journal of
Tropical Agriculture and Food Science 35.1 , 183, 2007.
[10] Nagle, Marcus, et al. "Non-Destructive Mango Quality Assessment Using
Image Processing: Inexpensive Innovation for the Fruit Handling
Industry." Conference on International Research on Food Security, Natural
Resource Management and Rural Development, 2012.
[11] Aibinu, Abiodun Musa, et al. "Automatic fruits identification system using
hybrid technique." Electronic Design, Test and Application (DELTA),
2011 Sixth IEEE International Symposium on. IEEE, 2011.
[12] Pujari, Jagadeesh Devdas, Rajesh Yakkundimath, Abdulmunaf Syedhusain
Byadgi. "Grading and classification of anthracnose fungal disease of fruits
based on statistical texture features." International Journal of Advanced
Science and Technology 52.1, 121-132, 2013.
[13] Medjahed, Seyyid Ahmed. "A Comparative Study of Feature Extraction
Methods in Images Classification." International Journal of Image,
Graphics and Signal Processing (IJIGSP) 7.3, 16, 2015.
[14] Mery, Domingo, Franco Pedreschi, and Alvaro Soto. "Automated design of
a computer vision system for visual food quality evaluation." Food and
Bioprocess Technology 6.8, 2093-2108, 2013.

You might also like