ISSN (Online) 2581-9429

IJARSCT
International Journal of Advanced Research in Science, Communication and Technology (IJARSCT)

Volume 2, Issue 4, May 2022

Impact Factor: 6.252

CIFAR-10 Image Classification with Convolutional

Neural Networks
Praneeta Handral, Ritika Kulkarni, Swapna MD, Nikhil Kumar
Students, Department of Computer Science and Engineering,
Alva’s Institute of Engineering and Technology, Mangalore, India

Abstract: In this project, we work on image classification of the CIFAR-10 dataset using supervised machine
learning techniques. The dataset consists of 60,000 32x32RGB images containing one of 10 object classes,
with 6000 images per class. We experiment with various learning algorithms including nearest neighbor
classifier, one-vs-all classification, Softmax classifier, two-layer fully connected artificial neural network
(ANN), deep convolutional neural network (CNN), and deep residual networks (ResNet). We use cross
validation by splitting the 50,000training data into49,000 training samples and 1,000 validation samples to
select the optimized hyper parameters for each parametric classifier. Among all methods, the 56-layer deep
residual network yields the best performance with a training accuracy above 99% and validation accuracy
of 93.6%.

Keywords: Image Classification; CIFAR-10; Supervised Machine Learning Algorithm; Deep Convolutional
Neural Network (CNN); Deep Residual Network (ResNet).

I. INTRODUCTION
Image classification is an active area of research and has been studied in common applications such as unmanned vehicles
and emergency robotics. In this paper, an integrated neural network (CNN)-based architecture is proposed using the Cifar
10 dataset, which has a total of 60,000 images [1]. These images are divided into training and testing sections, each with
50,000 and 10,000 images respectively. Although the CNN-based image classification methods presented herein are very
efficient, they require a large amount of memory. The purpose of our article is to perform image classification under limited
memory conditions. Such cases often occur in embedded systems. The proposed method achieves an accuracy of 85.9 while
requiring only 2 GB of GPU memory.

II. PROBLEM DEFINITION

The problem described is of importance throughout the field of automated equipment, which implements several diverse
classifiers to operate independently of human control. The most common automated applications can be seen in the
automotive and scientific industries, where scientists trying to automate vehicle operations or analyze medical scans, will
also come up with accurate diagnosis. Both the mentioned areas are very important for human because life depends on a
decision. The goal of automating devices and allowing them to take control of monotonous tasks is to make the most accurate
predictions.
Respond immediately. Knowledge of the fastest and most accurate image classification model can help multiple industries
implement the most appropriate algorithms. However, according to the no-free lunch theorem, each neural network
algorithm behaves differently on different datasets, so there is no single algorithm that solves all sorts of problems. For

Copyright to IJARSCT DOI: 10.48175/568 89

www.ijarsct.co.in
 ISSN (Online) 2581-9429
IJARSCT
International Journal of Advanced Research in Science, Communication and Technology (IJARSCT)

Volume 2, Issue 4, May 2022

Impact Factor: 6.252
example, the CNN model examined in this study should help classify a small number of classes that consist primarily of
small images. Attempts to resolve the issues presented are performed on the widely used CIFAR10 dataset. It is often used
to evaluate image classification algorithms. The dataset contains "60,000 32x32 color images in 10 layers, 6,000 images in
each layer" [3]. The full set of images is split into 2 sets of 50,000 images for the training set and the remaining 10,000
images for the test set. The training pack contains 5000 images for each class and the test pack contains 1000 representative
images for each class. The CIFAR10 dataset is a set of 10 classes divided into 6 animal categories (birds, cats, deer, dogs,
frogs, horses) and 4 vehicle types (airplanes, cars, boats, trucks). As the editors pointed out, the classes are mutually
exclusive and do not have the following overlapping similarities: B. Between the truck and the car. The CIFAR10 dataset
can be downloaded from the official website [5]. These files are available for download in a variety of Python, Matlab, and
binary versions suitable for C programs. After downloading the Python version (cifar10 python.tar.gz), the data will be
extracted and you will see the following file:
 batches.meta - includes a Python dictionary object that identifies the tag names of the 10 included classes.
 data_batch_1, data_batch_2,…, data_batch_5 - training data of 50,000 images divided into five files, 30 MB each.
 readme.html - HTML document that links to the official website of the dataset.
 test_batch - test data of 10,000 images in a 30MB file.
To view the actual images and use them, the files must decoded with a Python script.

III. PROPOSED WORK

The main object of the present study is to test and compare the performance of different CNN models performed with the
use of scripts written in the Python programming language. The small image classification pipeline used in this project is a
modern reconstruction, removing some of the components used in the standard CNN model. If successfully implemented,
the reinvented model will improve the classification rate. Test the deep neural network model applied on one of the famous
image classification datasets, that is the CIFAR10 dataset, because the results can then be compared with the rest of the
solutions announced.

Figure 1: Proposed block diagram

Step 1: We can download the CIFAR-10 dataset from reputable website.
Step 2: Before constructing a version we are able to decode the dataset to peer the real pictures and use them, the documents
must be decoded with the usage of a Python script.
Step 3: We can construct CNN version and educated that fashions on education dataset.
Step 4: We can take a look at the version on checking out dataset and examine the overall performance end result of each
the version and evaluate the overall performance.
Copyright to IJARSCT DOI: 10.48175/568 90
www.ijarsct.co.in
 ISSN (Online) 2581-9429
IJARSCT
International Journal of Advanced Research in Science, Communication and Technology (IJARSCT)

Volume 2, Issue 4, May 2022

Impact Factor: 6.252
IV. EXPERIMENTAL AND RESULT ANALYSIS
Running deep neural network models is very time consuming. All software is installed on Python, allowing Jupyter
Notebook to be used to run all the code. The Python programming language is implemented in a decoder file to import the
CIFAR10 dataset into a Jupyter notebook in a compatible format.
To run CNN models, Python requires the following packages:
 Tensorflow 1.7.0
 Keras 2.1.5
 Pickle (Python’s built-in package)
 NumPy 1.14.2
 Matplolib 2.2.2
The Keras library is a key component of the entire setup that initializes the CNN model [14]. This package allows you to
create a backend based on CPU or GPU components. This can change the execution time of the experiment. First, all Python
projects require you to import used packages, as shown in Figure 2. This makes it easier to enter the code. The main
component of the next setup is the Keras package running in Tensorflow-Open source machine learning framework
developed by Google that enables faster implementations of CNN networks in Python script.

Figure 2: Imports of packages

Second, the project requires importing the file "decoder.py" containing a set of functions written to decode the CIFAR10
dataset using the Pickle package and plot the labels and image data into the tables using the NumPy library. The code for
the entire decoder can be accessed from the link at the end of this chapter for further study.

Figure 3: Process of decoding and fetching data

Copyright to IJARSCT DOI: 10.48175/568 91
www.ijarsct.co.in
 ISSN (Online) 2581-9429
IJARSCT
International Journal of Advanced Research in Science, Communication and Technology (IJARSCT)

Volume 2, Issue 4, May 2022

Impact Factor: 6.252
The next step requires using the functions imported from the "decoder.py" file to load the layer name, check the number
of layers, and set the size of the input image, which is 32 x 32 pixels, and specify the channel number as three ( red, green
and blue). The three arrays of numbers are made up of values from 0 to 255, indicating the intensity of the pixels at that
point. With this information, the CNN can describe the probability that an object belongs to a particular class.
Another important operation is decryption and image recovery. Class labels use integer data type, while class use onehot
encoded vector. The CIFAR10 dataset was divided into two sets: 83% (50,000) images for training and the remaining 17%
(10,000) for testing, tested using “print()” shown in Figure 3.
The Keras library makes it very intuitive to create a model because you can use the model.add () function to define each
layer with a single line of code. The code used for the entire operation is self-explanatory, and after executing the function
shown in Figure 4, we use four 2D layers to summarize the CNN model. First, the model is initialized with a sequential
function that can build a linear stack of layers that are treated as a stack of objects, and each layer passes data to the next
layer. The first two "Conv2D (32, (3, 3)" scripts each initialize 32 convolution filters of 3x3 size, and then two more
"Conv2D (64, (3, 3)" scripts are more Use many filters.

Figure 4: Simple CNN model building

Afterwards, the model is trained on the training data. The ModelCheckpoint () method saves the subsequent optimal model
All ages. The compilation method described in the previous chapter specifies the loss function, optimizer, and model
evaluation metrics. Finally, the model fits the provided data to a batch size of 128. This is the number of samples per gradient
update. The model uses 100 iterations (epochs). After successful training, the model is evaluated and the accuracy and loss
are plotted on the Matplotlib graph. The IPython notebook contains some scripts for predicting the test image classes shown
in Figure 5.

Figure 5: Sample predictions

Copyright to IJARSCT DOI: 10.48175/568 92
www.ijarsct.co.in
 ISSN (Online) 2581-9429
IJARSCT
International Journal of Advanced Research in Science, Communication and Technology (IJARSCT)

Volume 2, Issue 4, May 2022

Impact Factor: 6.252
4.1 Improved CNN Model
Firstly, the pocket book calls for uploading some extra features from Keras and decoder packages. The manner of
uploading magnificence names in addition to fetching and deciphering the facts stays unchanged. Definition of the stepped
forward CNN version includes the maximum vital modifications proven in figure 6. As referred to withinside the
introduction, the stepped forward version replaces the max-pooling and dense feature with two- dimensional convolution
layers. The structure makes use of 9 layers with a specific quantity of convolution filters. In the end, the version makes use
of the operation of two-dimensional worldwide common pooling. After the shape definition, the version is constructed and
summarized.
The model is trained using the same parameters, where the only difference is the increased number of 350 epochs. and
the prediction result are shown in figure 6.
Evaluate the model

Figure 6: Improved Model prediction

4.2 Comparison
The accuracy, as well as running time of all the tested models, are presented in the following table

Figure 7: Comparison of models

The results show a 10% improvement between the simplest model and the most advanced model used in the test. The
CNN model after removing the maxpooling function and the density function improved the accuracy up to 87.94%; however,
rom three times as many epochs, uptime has increased for enhanced CNNs. Now we illustrate the model accuracy and
model loss of each of the tested CNN model shown in figure 8.

Figure 8: Simple CNN model accuracy and loss

Copyright to IJARSCT DOI: 10.48175/568 93
www.ijarsct.co.in
 ISSN (Online) 2581-9429
IJARSCT
International Journal of Advanced Research in Science, Communication and Technology (IJARSCT)

Volume 2, Issue 4, May 2022

Impact Factor: 6.252
From the first figure, it can be observed that the model stops improving its accuracy after about 60 epochs, while the loss
stabilizes after about 40 epochs. The second graph shown in Figure 9 of the advanced model shows the same situation as
the simple CNN model. Similarly, the model stops improving its accuracy after about 60 epochs, while the model loss starts
to increase slowly after about 150 epochs, resulting in overfitting of the data.

Figure 9: Improved CNN model accuracy and loss

V. CONCLUSION
The results obtained in this study are important because they suggest that the accuracy of the CNN model can be improved
simply by using a programming language to execute and modify traditional structures. One of the interesting conclusions is
the ratio of accuracy to delay. This is because the last model required the most computational power to achieve the highest
accuracy, while the most traditional CNN structure achieved the highest delay-to accuracy ratio. Based on existing
components, AI implementers need to determine if it is worth relying on a more robust model.

REFERENCES
[1]. Raniah Zaheer , Humera Shaziya “A Study of the Optimization Algorithms in Deep Learning” in IEEE 2019.
[2]. K. Simonyan and A. Zisserman, “Very deep convolutional networks for Large-Scale image recognition,” Sep. 2014.
[3]. Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep Learning. MIT Press, 2016.
[4]. https://www.ijcaonline.org/archives/volume176/nu mb er37/pandit-2020-ijca-920489.pdf
[5]. http://www.divaportal.org/smash/get/diva2:11111 44/F ULLTEXT02.pdf
[6]. https://scihub.hkvisa.net/10.1109/IAEAC47372.2019. 8997743
[7]. https://scihub.hkvisa.net/10.1109/AICCSA.2018.861 2 873
[8]. http://www.divaportal.org/smash/get/diva2:11111 44/F ULLTEXT02.pdf
[9]. https://www.ijert.org/rfid-based intelligent- busmanagement-and-monitoring system
[10]. https://scihub.hkvisa.net/10.1109/ICCASM.2010.5 62 0407
[11]. https://scihub.hkvisa.net/10.1109/IAEAC47372.2019 .8997743
[12]. https://scihub.hkvisa.net/10.1109/IAEAC47372.2019 .8997743
[13]. https://ieeexplore.ieee.org/document/8997743
[14]. http://www.divaportal.org/smash/get/diva2:1111144/FULLTEXT02.pdf

Copyright to IJARSCT DOI: 10.48175/568 94

www.ijarsct.co.in