ISSN 2278-3091

Volume 8, No.5, September - October 2019

International Journal of Advanced Trends in Computer Science and Engineering
Available Online at http://www.warse.org/IJATCSE/static/pdf/file/ijatcse78852019.pdf
https://doi.org/10.30534/ijatcse/2019/78852019

Classification of Mushroom Fungi Using Machine Learning Techniques

Mohammad Ashraf Ottom1, Noor Aldeen Alawad2, Khalid M. O. Nahar2

1
Department of Information Systems, Yarmouk University, Jordan
2
Department of Computer Sciences, Yarmouk University, Jordan

ABSTRACT the human immune system. Currently, the mushroom refers to the
process that performed by robot in food industry. This technique
Mushroom is one of the fungi types’ food that has the used to limit the features such as color. Recently, mushroom
most potent nutrients on the plant. Mushrooms have major system used specific characteristics that improve the selection
medical advantages such as killing cancer cells. This study process of mushrooms. Such system depends on analyzing and
aims to find the most appropriate technique for mushroom investigating the features in order to get better classification
classification, and mushroom will be classified into two based on the well-known features [3].
categories, poisonous and nonpoisonous. The proposed
approach will implement a different techniques and 1.1 Machine Learning
algorithms like neural network (NN), Support Vector
Machines (SVM), Decision Tree, and k Nearest Neighbors Machine learning (ML) relies under the umbrella of artificial
(KNN), on dataset of mushroom images, where the dataset intelligence [4], that allows computer systems to learn based on
contains images with background and without background. previous history, experience, examples, and data, it has been
The experimental results shown that the best technique for making great progress in many directions. Machine learning
classifying mushroom images is kNN with accuracy of 94% involving the study of computational learning and pattern
based on features extracted from images with real recognition theory in the artificial intelligence. In addition,
dimensions of mushroom types, and 87% based on features machine learning spots the light on the construction of techniques
extracted from images only. that can learn and make predictions on available data. In instance,
applications such as detection of network intruders or email
Keywords: Machine Learning, Mushroom Classification, filtering, optical character recognition (OCR), and computer
Supervised Learning.
vision [5].
1. INTRODUCTION
1.2 Algorithms and Techniques
Nowadays, there are different challenges to develop
systems that analyze a huge and complex data to make better In this study, we will use different machine learning
decisions. This study aims to find new approach working to algorithms and techniques for mushroom classification, some of
classify the mushrooms images based on different features them are listed below:
using the different techniques of Machine Learning (ML).
The purpose of classification process is to predict categorical  Neural Network (NN): is a distributed matrix structure, it
class labels or the target value [1], for example, feed-forward used in different applications, such as classifying data and
Artificial Neural Network (ANN), and the purpose of patterns, , predicting new cases or examples, and in pattern
classifier is to map data to predefined classes or groups [2]. recognition applications. NN simulates human biological
In the proposed approach, we used the training dataset that cells and human capability of thinking and learning [5][6].
contain the mushroom images to classify it into poisonous  Decision Tree: is one of the most popular classification
and nonpoisonous. Where our approach aims to classifies techniques in machine learning, where it used in decision
and predict for the class (groups) of mushrooms when submit support system. Decision aims to classify objects (instances)
the features of the mushrooms to different techniques of to find a track from the great parent node [7][8].
machine learning.  kNN: is an algorithm and classified under machine learning.
The kNN featured the low number of training parameters,
A mushroom is one of the fungi types’ food that has the where the computational complexity is not high, and the
most potent nutrients on the plant. Mushrooms have major performance is satisfactory [7].
advantages such as kill cancer cells, viruses and enhancing
2378
Mohammad Ashraf Ottom et al., International Journal of Advanced Trends in Computer Science and Engineering, 8(5),September - October 2019, 2378- 2385

2. LITERATURE REVIEW suggested method is depending on Euclidean distance measure. In

the suggested algorithm, the data set converted into numeric
There are different researches using of different techniques values. Then, the algorithm read the input data with normalizes
that are used for mushrooms classification. a Mushroom the numeric attributes to avoid the wide range of values. The
Diagnosis Assistance System (MDAS) was proposed by [3], experiment result showed that the suggested modified K-means
which involves three components of web application techniques faster compared to the existing algorithm.
(server), unified database and mobile phone application
(client) which is used on mobile phone devices. The Naive Al-mejibli and Hamad in [1] developed an application can
Bays and Decision Tree classifiers are used to determine the be applied on a mobile phone and web application named
mushroom types. Firstly, the suggested system chooses the Mushroom Diagnosis Assistance System, the purpose of this
most known mushroom attributes. Secondly, specify the application is to realize safety when gathering mushroom. They
mushroom type. The experiment results show that Decision used decision tree and naïve bays classifiers to group the
Tree classifier is better than Naïve Bays classifier in correct mushrooms types. They depended on the most famous mushroom
and incorrect classified instances, and error measurements. attributes to determine the mushroom type. This model has to
main phases: training phase and selection phase, to assign most
Kumar and others in [9] compared different active features in selection process and locate the final decision.
classification techniques that are used in data mining for The experimental results showed that decision tree was better
decision systems. A comparison take place among three than naive bays based on error measurements, correctly classified
decision trees algorithms represented by one statistical, one samples and incorrectly classified samples. The authors of [12]
artificial neural network, one support vector machines and analyzed a previous mushroom data set by using different data
one clustering algorithm. The suggested approach uses four mining techniques and Weka mining tool. They used nearest
datasets from several domains to test the predictive accuracy, neighbor classifier, covering algorithm to collect correct rules,
error rate, comprehensibility, classification index and unpruned decision tree and a voted perceptron algorithm. They
training time. The experimental results showed that Genetic reached from running the techniques on different groups by
Algorithm (GA) and support vector machines algorithms are stockholders that unpruned tree gives the best accuracy result and
better compared with the others in the predictive accuracy then it used on human-machine application based on web to
metric. In decision tree-based algorithms, QUEST algorithm produce interactive mushroom identification.
generates trees with smaller breadth and depth. In
conclusion, the GA based algorithm is the best algorithm that Chowdhury and S. Ojha in [13] identified a manner to
can be used for their decision support systems. distinguished several mushroom diseases using different data
mining classification methods. They used actual dataset gathered
Babu and others in [10] proposed a new application from mushroom farm by using data mining like Naïve Bayes,
domain that is used for SVM. The suggested approach uses RIDOR and SMO algorithms. They performed comparison based
the Support Vector Machine and Naïve Bayes algorithms for on a statistical way to detect popular symptoms for mushroom to
classification of mushrooms. The experiments results discover mushroom disease. They reached that naïve Bayes gives
showed that SVM is better compared to Naïve Bayer’s best result with comparisons to other classification techniques.
algorithm in term of accuracy. In conclusion, the SVM is an Beniwal and Das in [14] used data mining classification
efficient technique that can be used for application domain. techniques such as Zero, naïve Bayes and Bayes net to analyze
[2] used Multi-Layer Perception for Dataset training to create mushroom dataset that contain various kinds of mushrooms,
a model which is used to prediction of classifying. In the which are poisonous or not poisonous. They evaluated
experiment, only 8124 of dataset classification techniques by using accuracy, kappa statistic and
are used for training. The experiment result showed that the mean absolute error. They reached that Bayes net gives the lowest
best-hidden unit is 2, the best learning rates 0.6, the best mean absolute error and highest accuracy and then naïve Bayes.
activation function is sigmoid, the best moment rate is 0.2
and the best result of epoch is 300.

Onudu in [11] suggested modified K-means technique

based on the traditional k-mean algorithm to enhance the
clustering categorical dataset and solving the inherent
problem in the traditional clustering algorithm. The

2379
Mohammad Ashraf Ottom et al., International Journal of Advanced Trends in Computer Science and Engineering, 8(5),September - October 2019, 2378- 2385

3. METHODOLOGY There is a different information identify each type of mushroom

beside images, such as: family, location, dimensions, and
The aim of this study is to identify Mushroom images edibility. Figure 2 below shows an example of mushroom
and classify it into two categories (poisonous and images:
nonpoisonous) using machine learning techniques.

3.1 Research Phases

In this research our methodology consists of five phases:

the first phase is collecting dataset, second phase is
preprocessing the dataset, third phase is features extraction,
then machine learning model, and finally evaluation phase.
Figure 1 shows the research phases for the proposed
approach.

Figure 2: Dataset sample

3.3 Features Extracting

In this phase, we used Matlab for extracting all the features

from collected images in the raw dataset. Firstly, we extract the
Eigen features for each image after resizing it. Secondly, we take
the top 100 strongest. We use the dimension’s information that
available with each type in dataset and include this information
with feature matrix for the dataset, i.e. Cap diameter, stem tall
and diameter. Finally, we build the feature matrix, which contain
each of dimensions with the Eigen features with cap diameter,
stem tall and diameter, to build the Machine Learning (ML)
Model.

In the proposed approach, the ML model applied by

different techniques, such as: Neural Network (NN), Decision
Tree (DT), Support Vector Machine (SVM), and KNN. The best
results were for KNN with cross validation, where number of
folds equal 10, and accuracy 94%. Even though because of
difficulty to get the real measurements for mushrooms, we
decided to use extracted features from images only.

In order to try to enhance results, we find the width and

Figure 1: Research phases height for the shape of mushroom inside the pictures using
3.2 Collecting Dataset detecting edges in gray scale mode of images as shown in Figure
6. Then add these dimensions to the Eigen features. The
In the first step, we collected our dataset (raw dataset) of experiment results shown that KNN produced accuracy of 86%.
mushroom images from [15], where the collected dataset We attempted to extract different features from images to
consists of three categories (edible, inedible, and poisonous). enhance the results like histogram features. We applied the same
2380
Mohammad Ashraf Ottom et al., International Journal of Advanced Trends in Computer Science and Engineering, 8(5),September - October 2019, 2378- 2385

steps in previous experiment to calculate each of height and dataset. We applied NN, DT, SVM, and KNN algorithms on
width according to detect edges. We built new features matrix Orange 3 software, where the best results were for KNN with
for the dataset. The experiment results of histogram features 80% for the accuracy.
shown the accuracy reached to 87%. To optimize our results,
3.6 Machine learning model
we build an algorithm aims to extract more features we called
this group as parametric features, which are:
After feature extraction phase we use Orange3 & Knime to
 Local Contrast Normalization (LCN): used to build a machine learning model and apply different algorithms
contrast features within a feature map, as well as like SVM, Neural Network, Decision Tree and KNN. We use
across feature maps at the same spatial location, random sampling with training set size 66% but we didn’t get a
where it inspired by computational neuroscience good result due to the small number of the dataset instances
[16]. which were 380 instances. Therefore, we use cross validation
 Skewness, Standard deviation, and Kurtosis: are with 10 folds. After building the trained model we evaluate the
considering as concepts in the statistical meta- results in term of accuracy, f-measure, precision and recall. The
features, which are calculated by considering a confusion matrix is used to determine the percentage of wrongly
statistical concept, calculate this for all numeric classified instances. Figures 3 and 4 portrayed the training model
attributes and taking the mean [17]. in orange3 and Knime respectively.
 Entropy: is a concept with a complex history and has
been the subject of diverse reconstructions and
interpretations, where it’s defined as the average
amount of information produced by a stochastic
source of data [18].
 Mean: is useful in assessing expected losses and
benefits. For instance, in the proposed approach we
used mean to determine in features matrix,
especially for calculate the height and width for
images.
 Correlation: Correlation is one of the most widely
used, where the term "correlation" refers to a mutual
relationship or association between quantities [19].
 Homogeneity: is one of the broad categories of
distributed data mining, it refers to the process of
mining the same set of attributes over all the
participating nodes.
 Diameter: the real or virtual diameter of the length
of the mushroom stem tall.

3.4 Noise Reduction

Noise reduction is an important factor that influences

image quality [20], its working to reduce the errors of image
which it has problems. In the proposed approach, we will use Figure 3: Machine Learning model using Orange3
noise reduction to remove un-useful sections from original
images, such as background of images.

3.5 Features Extraction Images Without Background

In this phase, we used Matlab to extract Eigen features for

updated images (i.e. images without background), to build
features matrix again. Depending on the edges for the
mushroom images, we calculate the height and width for each
image, using detecting edges in gray scale mode to build new
2381
Mohammad Ashraf Ottom et al., International Journal of Advanced Trends in Computer Science and Engineering, 8(5),September - October 2019, 2378- 2385

Because the dimensions are not easy to measure when we have

only image for mushroom, we worked to extract virtual features
from images by calculating heights and widths for mushroom
images as it shown in Figure 6 below.

Figure 4: ML model using Knime

After using different tools to build machine learning model
we conclude that Knime is much faster than orange3, but
orange3 is user friendly and easier to use than Knime.
Figure 6: Width & height calculation
4. EXPERIMENT RESULTS
The results after calculating virtual dimensions from images and
After extracting all Eigen features from images, we add extracting Eigen features from images are shown in Figure 7
it to real dimensions, (cap diameter, stem tall). Figure 5 below. As we can see the best result obtained by KNN with
shows the results of accuracy for Eigen with real dimensions accuracy of 87%.
in different machine learning techniques.

Figure 7: Evaluation results for Eigen features

Figure 5: Evaluation results for Eigen features with real Next step we tried to extract more features like histogram
dimensions features, which applied to selected ML and shown the best
The experiment results shown KNN technique produced the accuracy gained by KNN technique with accuracy reached 87%.
highest accuracy (0.944) for Eigen with real dimensions. Figure 8 shows result for histogram features.

2382
Mohammad Ashraf Ottom et al., International Journal of Advanced Trends in Computer Science and Engineering, 8(5),September - October 2019, 2378- 2385

Figure 9: Evaluation results for parametric features

Figure 8: Evaluation results for histogram features

As we can see there is no enhancement on the results after
extracting these features. In order to gain better accuracy, we
We tried to add more features extracted from images which used another scenario by using Photoshop to remove
such as contrast, skewness, kurtosis, entropy, mean, standard backgrounds for images, and repeat all previous steps, but this
deviation, energy, correlation and homogeneity. We called scenario failed to give higher accuracy.
this group of features as parametric features. The
experimental results for this group of features is shown in Figure 10 shows the results for all techniques in proposed
Figure 9. approach (Neural Network, SVM, Decision Tree, and KNN) with
background images, while Figure 11 shows the results for the
same techniques without background images.

0.9
0.876 0.874
0.866 0.867 0.867
0.858
0.845 0.846 0.85 0.841
0.849
0.83 0.85

0.8
0.764
0.751
0.732 0.736 0.75

0.7

0.65
NN SVM Tree KNN

eigen eigen&parametric eigen&histogram&parametric histogram

Figure 10: Evaluation results for images with background

2383
Mohammad Ashraf Ottom et al., International Journal of Advanced Trends in Computer Science and Engineering, 8(5),September - October 2019, 2378- 2385

0.841 0.9
0.809 0.797 0.793
0.748 0.8
0.723
0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
histogram parametric eigen&histogram eigen eigen&histogram eigen&parametric

KNN Tree SVM NN

Figure 11: Evaluation results for images without background

5. CONCLUSION University College,” vol. 9, no. 2, pp. 103–113, 2017.

https://doi.org/10.29304/jqcm.2017.9.2.319
In the proposed approach, we used different algorithms to [2] M. Alameady, “Classifying Poisonous and Edible
get best results of mushroom classification, we implement each Mushrooms in the Agaricus,” International Journal of
of neural network (NN), SVM, Decision Tree, and KNN on Engineering Sciences & Research Technology, vol. 6,
no. 1, pp. 154–164, 2017.
different scenarios, with background and without background.
[3] R. LaBarge, “Distinguishing Poisonous from Edible
We extract different features from mushroom images like Wild Mushrooms,” 2008.
Eigen features, histogram features and parametric features. In [4] I. Kononenko, “Machine learning for medical
order to improve the results, we remove images background but diagnosis: history, state of the art and perspective,”
unfortunately this step failed to improve the result. Finally, the Artificial Intelligence in medicine, vol. 23, no. 1, pp.
experiment results show advantage for background images, 89–109, 2001.
especially when used KNN algorithm, and with Eigen features https://doi.org/10.1016/S0933-3657(01)00077-X
[5] L. Von Ahn, B. Maurer, C. McMillen, D. Abraham,
extraction and real dimensions of mushroom (i.e cup diameter,
and M. Blum, “recaptcha: Human-based character
stem tall and stem diameter) where accuracy reached to 0.944, recognition via web security measures,” Science, vol.
while the result after replacing real dimensions with virtual 321, no. 5895, pp. 1465–1468, 2008.
dimension (i.e. width and height of mushroom shape inside the https://doi.org/10.1126/science.1160379
images) is 87%. The highest value for KNN after removing [6] M. Tawarish and K. Satyanarayana, “A Review on
images background reached to 0.819 as a maximum value. Our Pricing Prediction on Stock Market by Different
Techniques in the Field of Data Mining and Genetic
future work we will try to extract some physical dimension
Algorithm,” International Journal of Advanced Trends
from mushroom images like cup diameters, stem tall, color and in Computer Science and Engineering, vol. 3, no. 23–
texture. Also, we will try to expand the dataset and use more 26, 2019.
images to improve classification process. https://doi.org/10.30534/ijatcse/2019/05812019
[7] N. Bhargava and G. Sharma, “Decision Tree Analysis
REFERENCES on J48 Algorithm for Data Mining,” International
Journal of Advanced Research in Decision Tree
[1] I. Al-Mejibli and D. Hamed Abd, “Mushroom Analysis on J48 Algorithm for Data Mining, vol. 3, no.
Diagnosis Assistance System Based on Machine 6, pp. 1114–1119, 2013.
Learning by Using Mobile Devices Intisar Shadeed Al- [8] A. Deshpande and R. Sharma, “Multilevel Ensemble
Mejibli University of Information Technology and Classifier using Normalized Feature based Intrusion
Communications Dhafar Hamed Abd Al-Maaref Detection System,” International Journal of Advanced
2384
Mohammad Ashraf Ottom et al., International Journal of Advanced Trends in Computer Science and Engineering, 8(5),September - October 2019, 2378- 2385

Trends in Computer Science and Engineering, vol. 8,

no. 3, pp. 874–878, 2019.
[9] P. Kumar, V. K. Sehgal, D. S. Chauhan, and others, “A
benchmark to select data mining based classification
algorithms for business intelligence and decision
support systems,” arXiv preprint arXiv:1210.3139,
2012.
https://doi.org/10.5121/ijdkp.2012.2503
[10] P. Babu, R. Thommandru, K. Swapna, and E. Nilima,
“Development of Mushroom Expert System Based on
SVM Classifier and Naive Bayes Classifier,”
International Journal of Computer Science and Mobile
Computing, vol. 3, no. 4, pp. 1328 –1335, 2014.
[11] F. E. Onuodu, “K-Modes Clustering Algorithm in
Solving Data Mining Problems for Mushroom
Dataset,” nternational Journal of Advanced Research
in Computer Science and Software Engineering, vol. 5,
no. 9, pp. 596–603, 2015.
[12] C. Eusebi, C. Gliga, D. John, and A. Maisonave, “Data
Mining on a Mushroom Database,” Proceedings of
Student-Faculty Research Day, pp. 1–9, 2008.
[13] D. Chowdhury and S. Ojha, “An Empirical Study on
Mushroom Disease Diagnosis : A Data Mining
Approach,” International Research Journal of
Engineering and Technology(IRJET), vol. 4, no. 1, pp.
529–534, 2017.
[14] S. Beniwal and B. Das, “Mushroom Classification
Using Data Mining Techniques,” International Journal
of Pharma and Bio Sciences, vol. 6, no. 1, pp. 1170–
1176, 2015.
[15] “Mushroom Dataset.”, Retrevided from
http://www.mushroom.world/.
[16] R. Socher, B. Huval, B. Bath, C. D. Manning, and A.
Y. Ng, “Convolutional-recursive deep learning for 3d
object classification,” in Advances in neural
information processing systems, 2012, pp. 656–664.
[17] G. Wang, Q. Song, H. Sun, X. Zhang, B. Xu, and Y.
Zhou, “A feature subset selection algorithm automatic
recommendation method,” Journal of Artificial
Intelligence Research, vol. 47, pp. 1–34, 2013.
https://doi.org/10.1613/jair.3831
[18] F. Flores Camacho, N. Ulloa Lugo, and H. Covarrubias
Martínez, “The concept of entropy, from its origins to
teachers,” EDUCATION Revista Mexicana de Física E,
vol. 61, no. December, pp. 69–80, 2015.
[19] R. Socher and B. Huval, “Convolutional-recursive deep
learning for 3D object classification,” Advances in
Neural …, no. i, pp. 1–9, 2012.
[20] C. Chang-yanab, Z. Ji-xian, and L. Zheng-jun, “Study
on methods of noise reduction in a stripped image,”
The International Archives of the Photogrammetry,
Remote Sensing and Spatial Information Sciences, vol.
XXXVII. Pa, no. 1, pp. 2–5, 2008.

2385