IRJET - Effective Workflow for High-Performance Recognition of Fruits using Machine Learning Approaches

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 07 Issue: 02 | Feb 2020 www.irjet.net p-ISSN: 2395-0072
© 2020, IRJET | Impact Factor value: 7.34 | ISO 9001:2008 Certified Journal | Page 1516
Effective Workflow for High-Performance Recognition of Fruits using
Machine Learning Approaches
Abdul Kader1, Samiha Sharif2, Pranta Bhowmick3, Fahmida Haque Mim4, Azmain Yakin Srizon5
1,2,3,4Dept. of CSE, Bangladesh Army University of Engineering & Technology, Bangladesh
5azmainsrizon@gmail.com, Lecturer, Dept. of CSE, Bangladesh Army University of Engineering & Technology,
Bangladesh
---------------------------------------------------------------------***----------------------------------------------------------------------
Abstract - Recognition system is an essential combat of
computer science. Fruit recognition is one ofthem. Researcher
have been roaming around this area for a decade now.
Previously, many traditional machinelearningtechniques and
deep learning methods have been employed for successful
recognition of fruits and in some cases, a high accuracy have
been achieved. However, in our study we showed that our
proposed working diagram outperformed all the previous
studies. In this paper, we started with Fruit-360datasetof109
distinct classes and applied three feature extraction methods
(hu moments, haralick texture and color histogram). After
that we applied several machine learning approaches
(Decision Tree, K-Nearest Neighbors, Linear Discriminant
Analysis, Logistic Regression, NaïveBayes, RandomForestand
Support Vector Machine) to train the models. Finally, the test
results were calculated and K-Nearest Neighbor along with
Random Forest classifiers produced best results with a false
positive rate of 0%, hence, achieving a high accuracy.
Key Words: Image Recognition, HU Moments, Haralick
Texture, Color Histogram, Decision Tree, K-Nearest
Neighbors, Naïve Bayes, Linear Discriminant Analysis,
Logistic Regression, Random Forest, Support Vector
Machine, ROC Curve Analysis
1. INTRODUCTION
Recognition practice has risen as a ‘grand challenge' for
computer vision, with the longer-term intention of being
capable to produce a near-human level of recognition for
tens of thousands of classes beneath a broad diversity of
circumstanceslikevisual perception,soundrealization,voice
identification, handwriting verification, text recognitionand
so on [1]. The face identification system automatically
identifies the input face pictures from digital image
processing. Text classification is practiced to recognize the
different types of text for example spam or non-spam email
[2]. However, in this research, our main purpose was the
recognition of fruits.
Color and texture are the primary aspectsofnatural pictures
and perform an essential part in visual understanding. It is
frequently beneficial to explain monochrome strain by
developing contrast or detachment. The manner of color
categorization includes the extraction of valuable
information concerning the spectral characteristicsofobject
exteriors and determining the most suitable match from a
set of associated records or class models to perform the
recognition task [3]. The texture is one of the most dynamic
subjects in machine intelligence and pattern investigation
ever since the 1950s which sought to distinguish different
decorations of images by removing the dependency of
intensity between pixels and their adjacent pixels [4] or by
capturing the variance of intensity across pixels.
Previously, several studies have been performed for the
successful recognition of fruits. A recent study on fruit
recognition from 38,409 images of 60 fruits using deep
learning has achieved an overall accuracy of 96.3% [5]. A
framework called Pure-CNN achieved a classification
accuracy of 98.88% for 81 categories of fruit-360 data
recently [6]. With the help of the K-Nearest algorithm
technique, a research study achievedanaccuracyof95%[7].
Another research suggested an overall accuracy of 98.33%
using neural classifiers [8]. However, in our research, we
studied a total of 109 fruit classeswith74,572images.Inthis
study, we’ve shown that our proposed working model has
outperformed all previous studies.
2. MATERIALS & METHODS
This section starts with the dataset description. After that,
the proposed model has been described. Next, several
machine learning techniques have been described briefly.
Finally, performance measurement has been discussed in
terms of different performance measurements.
2.1 Dataset Description
The fruit-360 dataset is one of the most traditional datasets
that includes 74,572 picturesof109assortedfruits[9].Fruits
were located in the shaft of a low-speed motor of 3 rpmanda
small documentary of 20 secondswasvideotaped.ALogitech
C920camera was utilized forrecording the outcomes.Thisis
one of the most trustworthy webcams available. Behind the
fruits, a white layer of paper was installed as a background.
Nevertheless, due to the fluctuations in the lighting
circumstances, the background was not consistent and the
dataset creators composed a dedicated algorithm that
extracts the fruit from the background. For further
processing, the dataset was divided into train and test set.
The train set contained 90% of total data having a total of
67,114 images. The test set had a total of 7,458 images.

2.2 Proposed Recognition Workflow
Firstly, the whole dataset was taken as input and several
feature extraction techniques had been performed i.e., Hu
Moments, Haralick Texture and Color Histogram. Secondly,
the features and labels of the images were saved in a numpy
array for further processing [10]. Thirdly, the dataset was
split into train and test set. The train set contained 90% of
the data where the test set contained 10% of the whole
dataset. After that, we applied several machine learning
techniques (Decision Tree, K-Nearest Neighbors, Linear
Discriminant Analysis, Logistic Regression, Naïve Ba yes,
Random Forest, and Support Vector Machine) to train the
model. With the help of 10-fold cross-validation, we
validated the overall accuracy of the trained model. Finally,
the models were applied to the test set for further
performance analysis. Our proposed working diagram has
been shown in Figure-1.
Figure-1: Proposed workflow diagram
2.3 Feature Extraction
Feature extraction starts from an initial setofmeasureddata
and builds derived values intended to be informative and
non-redundant, facilitating the subsequent learning and
generalization steps, and in some cases leading to better
human interpretations. In this study, we’ve used three
feature extraction techniques: HU Moments, Haralick
Texture and Color Histogram.
2.3.1 HU Moments
Hu Moments are a collection of7estimatesdeterminedusing
central moments that are invariant to image alterations.The
first 6 moments have been determined to be invariant to
translation, scale, and rotation, and reflection while the 7th
moment’s sign varies for image reflection [11].
2.3.2 Haralick Texture
Texture defines the persistence of decorations and colors in
an object/image such as woods, water, rocks, sands, etc. To
distinguish targets in an image based on texture, one has to
examine the uniform spread of decorations and colorsinthe
objects covering. Rough-Smooth, Hard-Soft, Fine-Coarseare
some of the texture combinations one can imagine of,
although there are several such combinations. Haralick
texture is applied to quantify an image based on texture. It
was developed by Haralick in 1973 [12]. The primary
thought associated with measuring Haralick Texture
characteristics is the Gray Level Co-occurrence Matrix or
GLCM.
2.3.3 Color Histogram
A histogram is recognized as a diagram or design which is
related to the regularity of pixels in a grayscale picture with
pixel values differing from 0 to 255. A grayscale photograph
is an image in which the value of an individual pixel is a
unique sample, that is, it brings only depth information
where pixel value differs from 0 to 255. Pictures of this kind,
also recognized as black-and-white, are formed solely of
shades of gray, ranging from black at the lowest intensity to
white at the strongest where a pixel can be viewed as a point
in an image.
2.4 Machine Learning Techniques
Several machine learning techniques were applied in this
study. All these approaches have been described in this
section.
2.4.1 Decision Tree
Decision Tree algorithm is a sorting algorithm for
constructing a decision [13]. This makes occasionally a
variance which means that in CART the conclusions on how
to split values based on an attribute are delayed.
2.4.2 K-Nearest Neighbors
The K-Nearest Neighbors algorithm is a simple, easy-to-
implement exercised machinelearningalgorithmthatcan be
utilized to solve both classification and regression matter
[14]. The algorithm is handy, humble and informal to
implement. There’s no requirement to form a model, tune
numerous parameters, or make additional assumptions.
2.4.3 Linear Discriminant Analysis
Linear Discriminant Analysis or LDA is a dimensionality
shortening procedure [15]. It isoperatedasa pre-processing
stage in Machine Learning and applications of pattern
grouping. The purpose of LDA is to project the characteristic
in higher dimensional place onto a lower-dimensional place
to avoid the damnation of dimensionality and also reduce
support and dimensional spending.
2.4.4 Logistic regression
Logistic regression is an adjustment algorithm employed to
produce measurements to a discrete set of groups [16].
Opposite linear regression that outputs continuous number

worth, logisticregressionchangesitsoutputapproaching the
logistic sigmoid function to proceed back a possible value
which can then be outlined to two or more discrete groups.
2.4.5 Naïve Bayes
Naïve Bayes is a probabilistic machine learning algorithm
that can be applied in a wide-ranging varietyofclassification
tasks [17]. It is very simple, relaxed to implement and fast,
required fewer training data. If the Naïve Bayes con ditional
individuality statement holds, then it will appear quicker
than discriminative models like logistic regression.
2.4.6 Random Forest
Random Forest is grounded on decision trees [18]. In
machine learning, decision trees are a fetch for creating
predictive models. It is practiced to resolve equally
classification as well as regressionproblems.RandomForest
can habitually handle lost values. It is fairly less impacted by
noise. Random Forest algorithm is right stable. Evenifa new
data point is announced in the dataset, the total algorithm is
not affected much since the new data may affect one tree.
2.4.7 Support Vector Machine
Support Vector Machine or SVM algorithm is an easy yet
strong supervised machine learning algorithm that can be
exercised for constructing both regressionandarrangement
models [19]. SVM algorithm can act really well with both
linearly dividable and non-linearly separable datasets.
2.5 Performance Measurement
For measuring the performance of the classifiers, we’ve
applied the measurements of sensitivity, specificity,
precision, f-score, false-positive rate, and overall accuracy.
All of these phenomena have been described below. Here,
TP, TN, FP, and FN correspond to truepositive,truenegative,
false positive and false negative respectively. ROC curve
analysis was also performed in our study.
2.5.1 Sensitivity
Sensitivity is described as the probability of accurately
recognizing some conditions. For example, the sensitivity
might be employed in a medical investigationto reportthata
particular test has an 80% probabilityofidentifyinganabolic
steroid utilized by an acrobat. Sensitivity is calculated with
the following formula:
Sensitivity = TP/(TP+FN)
2.5.2 Specificity
Specificity contains a test’s strength to precisely produce a
negative result for characters who don’t hold the condition
that’s being tested for. A high-specificity test will perfectly
rule out nearly everyone who doesn’thavethecondition and
won’t produce many false-positive results. Specificity is
measured with the following formula:
Specificity = TN/(TN+FP)
2.5.3 Precision
Precision points to how familiar estimations from separate
samples are to each other. For example,thestandard erroris
an example of precision. When the standard error is little,
estimations from different sampleswill bealikeinvalue;and
vice versa. Precision is measured as follows:
Precision = TP/(TP+FP)
2.5.4 F-Score
The F score is determined as the weightedharmonicmeanof
the test's precision and recall. This score is measured
according to the precision and recall of a test practiced into
account. F-Score is estimated with the help of the following
formula:
F-Score = 2TP/(2TP+FP+FN)
2.5.5 False Positive Rate
In statistics, when conducting various comparisons, a false
positive ratio is the probability of incorrectly discarding the
null hypothesis for a distinct test. The false-positive rate is
determined as the ratio between the number of negative
results incorrectly classified as positive (false positives)and
the total amount of actual negative results.
False Positive Rate = FP/(FP+TN)
2.5.6 Overall Accuracy
Overall accuracy is the possibility that a sample will be
accurately matched by a test; that is, the total of the true
positives and true negatives divided by the total number of
individuals examined that is the sum of true positive, true
negative, false positive and false negative. However, overall
accuracy doesn’t show the actual performance as sensitivity
and specificity may differ despite having a higher accuracy.
Overall accuracy can be estimated as follows:
Overall Accuracy = (TP+TN)/(TP+TN+FP+FN)

Table-1: Sensitivity, specificity, precision, f-score, false-positive rate (FPR) and accuracy measurement for Decision Tree
(DT), K-Nearest Neighbors (KNN), Linear Discriminant Analysis (LDA), Logistic Regression (LR), Naïve Bayes (NB),
Random Forest (RF) and Support Vector Machine (SVM) where the range of the measurements is [0,1]
Approach Sensitivity Specificity Precision F-Score FPR Accuracy
DT 0.99990 0.99038 0.99990 0.99990 0.00961 0.99021
KNN 1.00000 1.00000 1.00000 1.00000 0.00000 1.00000
LDA 0.99985 0.98418 0.99985 0.99985 0.01581 0.98417
LR 0.99997 0.99757 0.99997 0.99997 0.00242 0.99758
NB 0.99967 0.96984 0.99967 0.99967 0.03015 0.96486
RF 1.00000 1.00000 1.00000 1.00000 0.00000 1.00000
SVM 0.99817 0.82869 0.99816 0.99815 0.13461 0.80155
2.5.7 ROC Curve Analysis
ROC curves are regularly utilized to present graphically the
association/trade-off among sensitivity and specificity for
every reasonable cut-off for a test or a sequenceoftests[20].
Besides, the section under the ROC curve provides an
opinion about the advantage of practicing the test in
question. As the field under a ROC curve is a measure of the
usefulness of a test in general, where a larger area indicates
a more valuable test, the areas under ROC curves are
adopted to distinguish the usefulness of tests. There are two
types of ROC curves in terms of multi-class ROC curve
analysis: macro-average ROC curve and micro-average ROC
curve. A macro-average ROC curve measures the metric
freely for each class and then takes the average, whereas a
micro-average ROC curve aggregates the participation of all
classes to calculate the average metric. Although in a multi-
class classification setup, micro-average is favored, in our
study, we took both of them under consideration.
3. Experimental Analysis
The experiment started by taking the fruit-360 dataset. At
first, three feature extraction processes were applied on the
whole dataset: hu moments, haralick texture and color
histogram. All these results were then stored in numpyarray
for further processing. Next, the dataset was split into train
and test sets. The train set contained 90% of the data and the
test set contained the rest 10%. After that, several machine
learning techniques were performed on the train set for
fitting the models. Scikit learn was used for training these
models. While implementing the models, all the parameters
were set to default. 10-fold cross-validation was used for
validating the train set results. After the validation, trained
models were performed on test data. Sensitivity, specificity,
precision, f-score and false positive rate were measured for
each class and the mean of these measurements was
considered as the overall sensitivity, specificity, precision, f-
score, and false-positiveraterespectivelyforaspecificmodel.
Overall accuracy was also measured for each of the models.
Table-1 shows all the measurements for all the models.
Figure-2 shows the comparison of accuracy for the machine
learning classifiers: decisiontree, k-nearestneighbors,linear
discriminant analysis, logistic regression, naïve bayes,
random forest and support vector machine.
Figure-2: Comparison of accuracy among all seven
machine learning classifiers
From Table-1 and Figure-2, it can be assured that in our
proposed model K-Nearest Neighbor and Random Forest
algorithms outperforms other methods. Moreover, our
results also outperformed previous studies. However, for
further evidence, the ROC curve analysis was performed as
well. We generated both macro-average and micro-average
ROC curves. Figure-3 shows the macro-average ROC curve
analysis for all seven classifiers.

Figure-3: Macro-average ROC curve analysis for seven
classifiers
In our research, we also found out the micro-average ROC
curve. Figure-4 shows the micro-averageROCcurveanalysis
for all seven classifiers.
Figure-4: Micro-average ROC curve analysis for seven
classifiers
Hence, from all these experimental results, we concluded
that our proposed model outperformed all other studies of
fruit recognition for fruit-360 datasets for the K-Nearest
Neighbors and Random Forest classifiers.
4. CONCLUSION
Researches for fruit recognition on fruit-360 dataset has
been circulating through the research community for a long
time now. So far, many studies have been performed and a
higher accuracy has been achieved on some of the cases. But
in our study, we proposed a working diagram that produced
an extremely well accuracy for K-Nearest and Random
Forest classifiers. Fruit-360 is updated regularly. Hence,
there’s always a chance for improvement in this sector. But
for the time being our proposed method outperformed all
previous studies.
REFERENCES
[1] Duda RO, Hart PE, Stork DG. Pattern classification. John
Wiley & Sons; 2012 Nov 9.
[2] Bruce V, Young A. Understanding face recognition.
British journal of psychology. 1986 Aug;77(3):305-27.
[3] Kartikeyan B, Sarkar A. An identificationapproachfor2-
D autoregressive models in describing textures. CVGIP:
Graphical Models and Image Processing. 1991 Mar
1;53(2):121-31.
[4] Sahin, F., 1997. A radial basis function approach to a
color image classification problem in a real time
industrial application (Doctoral dissertation, Virginia
Tech).
[5] Mureşan H, Oltean M. Fruit recognition from images
using deep learning. Acta Universitatis Sapientiae,
Informatica. 2018 Aug 1;10(1):26-42.
[6] Kausar, A., Sharif, M., Park, J. and Shin, D.R., 2018,
December. Pure-CNN: A Framework for Fruit Images
Classification. In 2018 International Conference on
Computational Science and Computational Intelligence
(CSCI) (pp. 404-408). IEEE.
[7] Ninawe P, Pandey S. A completion on fruit recognition
system using k-nearest neighbors algorithm.
International Journal ofAdvancedResearchinComputer
Engineering & Technology (IJARCET). 2014
Jul;3(7):2352-6.
[8] Priya MM, Sujatha N. RECOGNITION OF FRUITS USING
MULTIPLE FEATURES WITH NEURAL CLASSIFIERS.
Asian Journal of Computer Science And Information
Technology. 2017 Jul 3;7:37-9.
[9] Mureşan H, Oltean M. Fruit recognition from images
using deep learning. Acta Universitatis Sapientiae,
Informatica. 2018 Aug 1;10(1):26-42.
[10] Walt SV, Colbert SC, Varoquaux G. The NumPy array: a
structure for efficient numerical computation.
Computing in Science & Engineering. 2011 Mar;
13(2):22-30.
[11] Huang Z, Leng J. Analysis of Hu's moment invariants on
image scaling and rotation. In2010 2nd International
Conference on Computer Engineering and Technology
2010 Apr 16 (Vol. 7, pp. V7-476). IEEE
[12] Haralick RM, Shanmugam K, Dinstein IH. Textural
features for image classification. IEEE Transactions on
systems, man, and cybernetics. 1973 Nov(6):610-21.
[13] Safavian SR, Landgrebe D. A survey of decision tree
classifier methodology. IEEE transactions on systems,
man, and cybernetics. 1991 May; 21(3):660-74.

[14] Dudani SA. The distance-weighted k-nearest-neighbor
rule. IEEE Transactions on Systems, Man, and
Cybernetics. 1976 Apr(4):325-7.
[15] Balakrishnama S, Ganapathiraju A. Linear discriminant
analysis-a brief tutorial. Institute for Signal and
information Processing. 1998 Mar 2;18:1-8.
[16] Kleinbaum DG, Dietz K, Gail M, Klein M, Klein M. Logistic
regression. New York: Springer-Verlag; 2002 Aug.
[17] Zhang H. The optimality of naiveBayes.AA.2004;1(2):3.
[18] Liaw A, Wiener M. Classification and regression by
randomForest. R news. 2002 Dec 3;2(3):18-22.
[19] Cortes C, Vapnik V. Support-vector networks. Machine
learning. 1995 Sep 1; 20(3):273-97.
[20] Fawcett T. An introduction to ROC analysis. Pattern
recognition letters. 2006 Jun 1; 27(8):861-74.

IRJET - Effective Workflow for High-Performance Recognition of Fruits using Machine Learning Approaches

Recommended

Recommended

More Related Content

What's hot

What's hot (20)

Similar to IRJET - Effective Workflow for High-Performance Recognition of Fruits using Machine Learning Approaches

Similar to IRJET - Effective Workflow for High-Performance Recognition of Fruits using Machine Learning Approaches (20)

More from IRJET Journal

More from IRJET Journal (20)

Recently uploaded

Recently uploaded (20)

IRJET - Effective Workflow for High-Performance Recognition of Fruits using Machine Learning Approaches