Integration of Big Data Processing Tools and Neural
Networks for Image Classification
Nikita E. Kosykh Anatoly D. Khomonenko
Emperor Alexander I St. Petersburg Emperor Alexander I St. Petersburg State
State Transport University, Transport University,
Saint Petersburg, Russia Saint Petersburg, Russia
nikitosagi@mail.ru khomon@mail.ru
Alexander P. Bochkov Anatoly V. Kikot
Peter the Great St. Petersburg Emperor Alexander I St. Petersburg State
Polytechnic University, Transport University,
Saint Petersburg, Russia Saint Petersburg, Russia
kostpea@mail.ru a.v.kikot@yandex.ru
Deep learning of neural networks is a very popular
topic these days, especially in computer vision
applications. Until 2010, there was no
database[Liu16] sufficiently complete and
Abstract voluminous in order to train a neural network in a
high-quality way to solve certain problems related
The issues of joint use of tools for to image recognition. Existing solutions had a high
processing big data in solving problems of level of inaccuracy. With the emergence of open
artificial intelligence are becoming
databases, such as Kaggle (big data bank) and
increasingly important. The article
ImageNet (image bank), it is possible to train
discusses the task of optimizing the
neural networks and make practically error-free
parameters of neural networks used for
image recognition using Matlab and decisions. The article discusses the task of
Hadoop systems, as well as the MNIST optimizing the parameters of neural networks used
(Modified National Institute of Standards for image recognition using Matlab and Hadoop
and Technology) database is a voluminous systems.
database of handwritten number samples. Justification of the choice of programming
The results of calculating the optimal environment. The Matlab 2016a[Kra18] chosen as
number of neural network layers for an integrated programming environment for
solving the classification problem of working with big data. It supports some tools for
images presented. We study the issues of solving problems, namely:
evaluating the accuracy of image • Variables stored on the hard disk. Using the
classification depending on the number of matfile function, you can access MATLAB
network neurons, choosing the optimal variables directly from the MAT file on disk,
network training algorithm, and evaluating without loading the entire variable into
the effect of parallelization[Sha09] using
memory. This will allow processing of large
MATLAB Distributed Computing Server in
data sets that do not fit in memory.
the process of training a neural network on
computing performance. • Datastore. The datastore function allows
you to access data that does not fit in
memory. This may include data from files,
1 Introduction sets or tables.
1The issues of joint use of tools for processing • Parallel computing. Parallel Computing
Toolbox implements parallel loops to run
big data[Bah15] in solving problems of artificial
MATLAB code on multi-core architectures.
intelligence are becoming increasingly important
Machine learning. Machine learning used to
Copyright c by the paper's authors. Use permitted under Creative
develop predictive models for big data. The whole
Commons License Attribution 4.0 International (CC BY 4.0). range of machine learning algorithms is contained
In: A. Khomonenko, B. Sokolov, K. Ivanova (eds.): Selected in the Statistic Toolbox and Neural
Papers of the Models and Methods of Information Systems Research
Workshop, St. Petersburg, Russia, 4-5 Dec. 2019, published at
Network[Pao89] Toolbox modules.
http://ceur-ws.org.
52
• hadoop. Using the MapReduce[Gha15] and To display an image, you need to create a
DataStore functionality built into MATLAB, you can graphic context (window) using the function figure.
develop algorithms on a personal computer and As parameters, you can pass properties for a
run them on Hadoop. You can request a piece of graphic context to a function. Initially handwritten
data using the datastore function, and then using numbers are displayed in colors that are presented
the Distributed Computer Server, run the in a color palette of green and blue shades (Fig. 1).
algorithms within the Hadoop MapReduce
environment on the complete set of data.
2 MNIST Dataset of Digit Patterns
The MNIST (Modified National Institute of
Standards and Technology) database is a
voluminous database of handwritten number
samples. The database is a standard proposed by
the US National Institute of Standards and
Technology for the purpose of calibrating and
comparing image recognition methods using
machine learning, primarily based on neural
networks.
Neural networks tend to learn better from Figure. 1: Display the default set
specific examples. In the development of For a better perception, we use the colormap ()
neural[Nov15] network for pattern function, which sets the color map of the final
recognition, we use the popular MNIST image.
handwritten data set (Hadoop, 2016). Kaggle's Colormap (gray) – sets a linear palette in shades
open portal for distributing big data [Peh19] of gray.
uses this very kit in the Digit Recognizer The following code fragment reads the first 16
training contest. The set contains the following lines of the data set, which are 16 digits. Converts
components: rows into matrices and displays the resulting
a) trainSet.csv – training data; images on the screen, while the numbers on the top
b) testSet.csv – test data for show the object class numbers.
presentation.
First you need to load the training data figure
colormap(gray)
into MATLAB (MATLAB, 2018). For this we use
for i = 1:25
the built-in function csvread. subplot(5,5,i)
matrix = reshape(trainSet(i, 2:end),
M = csvread (filename, R1, C1) [28,28])'
trainSet = csvread (trainSet.csv, 1,0) imagesc(matrix)
title(num2str(tr(i, 1)))
testSet = csvread (testSet.csv, 1,0)
end
trainSet = csvread ('trainSet.csv', 1,0)
subSet = csvread ('testSet.csv', 1, 0);
After the above code fragment executed, an
image of 16 handwritten numbers will appear on
The first column in the trainset set is a label the screen, placed on a single graphic screen in a
that shows the correct number for each sample in black and white palette (Fig. 2).
the data set, and each line is a sample. In the
remaining columns, the row is an image of the
handwritten digit 28x28, but all the pixels are
placed in one row, and not in the original
rectangular shape. To render the numbers, we need
to rebuild the rows into 28x28 matrices. To do this,
you can use the Reshape function, with the
exception that you need to transpose the matrix,
because the Reshape function works in columns
and not line by line.
Figure. 2: Image handwritten numbers
53
To build the network[Kho15], we will use D. For the input, choose XtrainSet and for
the tool for pattern recognition ‘nprtool’ from classes, YtrainSet.
the Neural Network Toolbox module. E. After installing the arrays, go to the
“Data for verification” section. In this case, you
3 Data preparation can leave the default values. This will divide
The input tool expects two sets of data: the data in the ratio of 70-15-15 into sets for
training, testing and testing.
a) input – a numeric matrix, each column F. In the network architecture section,
of which represents samples and rows. change the value of hidden layers to 100 and
These are scanned images of go on.
handwritten numbers; G. For training, go to the Train Network
b) vectorLabels – matrix-row binary form section and click “Train” to start training. Upon
of 0 and 1, which is mapped to specific completion of the operation, a window with
labels that represent the image. It is learning results will open. Numerical
also called a dummy variable. Neural indicators can viewed as graphs and charts.
Network ToolBox also expects tags to H. On the last tab, you will prompted to
be stored in columns and not in rows. save the learning script for the model you just
Class labels range from 0 to 9, as there are created. (eg NeuralNetworkSctipt.m).
exactly so many single-digit numbers. To solve Below is a diagram of a model of an
this problem, we will use “10” instead of “0”, artificial neural network (Figure 3.), which was
because MATLAB indexing starts from 1. created using the pattern recognition tool. It
The trainSet dataset stores sample images has 784 input neurons, 100 hidden layer
in rows, not columns, so the entire dataset neurons and 10 output word neurons (which is
needs to transpose. The test set of data does equal to the number of classes for prediction).
not come out in advance in the MNIST set. To
form it, we will hold 1/3 of each data set
(training, test).
n = size(trainSet, 1);
labels = trainSet(:,1);
labels(labels == 0) = 10; Figure. 3: Neural network model with
labelsd = dummyvar(labels); parameters
inputs = trainSet(:,2:end)
inputs = inputs'; % transposition of the The model trained by adjusting the weights
set of predictors for correct results. W in the diagram denotes
labels = labels'; weights and b – displacement neuron, which
labelsd = labelsd'; are part of individual neurons. Separate
rng (1);
c = cvpartition (n, 'Holdout', n / 3); neurons in the hidden layer are as follows: 784
XtrainSet = inputs (:, training (c)); input neurons and correspondingly the same
YtrainSet = labelsd (:, training (c)); weights, 1 offset unit and 10 activation
XtestSet = inputs (:, test (c)); outputs.
YtestSet = labels (test (c));
YtestSetd = labelsd (:, test (c));
5 Data Visualization
Explanation: The function c = cvpartition For begin you should look inside the structure
(n, ‘Holdout, p) randomly creates a section for of the NeuralNetworkSctipt.m file. There you
validating validation during observations. This can see the generated variables, such as IW1_1
section divides the array into tutorials and a and x1_step1_keep, which are weights vectors
test set. Parameter p must be a scalar. When 0 obtained in the process of training the