Convolutional Neural Network (CNN) for Image Detection and Recognition

Abstract- Deep Learning algorithms are designed in such
a way that they mimic the function of the human cerebral
cortex. These algorithms are representations of deep neural
networks i.e. neural networks with many hidden layers.
Convolutional neural networks are deep learning algorithms
that can train large datasets with millions of parameters, in
form of 2D images as input and convolve it with filters to
produce the desired outputs. In this article, CNN models are
built to evaluate its performance on image recognition and
detection datasets. The algorithm is implemented on MNIST
and CIFAR-10 dataset and its performance are evaluated.
The accuracy of models on MNIST is 99.6 %, CIFAR-10 is
using real-time data augmentation and dropout on CPU
unit.
Keywords- Deep Learning, Handwritten digit Recognition,
Object Detection, Convolutional neural networks, MNIST,
CIFAR-10, Dropout, Overfitting, Data Augmentation, Relu
I INTRODUCTION
Image Recognition and detection is a classic machine
learning problem. It is a very challenging task to detect an
object or to recognize an image from a digital image or a
video. Image Recognition has application in the various
field of computer vision, some of which include facial
recognition, biometric systems, self-driving cars, emotion
detection, image restoration, robotics and many more[1].
Deep Learning algorithms have achieved great progress in
the field of computer vision. Deep Learning is an
implementation of the artificial neural networks with
multiple hidden layers to mimic the functions of the
human cerebral cortex. The layers of deep neural network
extract multiple features and hence provide multiple
levels of abstraction. As compared to shallow networks,
this cannot extract or work on multiple features.
Convolutional neural networks is a powerful deep
learning algorithm capable of dealing with millions of
parameters and saving the computational cost by inputting
a 2D image and convolving it with filters/kernel and
producing output volumes.
The MNIST dataset is a dataset containing
handwritten digits and tests the performance of a
classification algorithm. Handwritten digit recognition has
many applications such as OCR (optical character
recognition), signature verification, interpretation and
manipulation of texts and many more[2,3]. Handwritten
digit recognition is an image classification and
recognition problem and there have been recent
advancements in this field [4]. Another dataset is CIFAR-
10 which is an object detection datasets that classifies the
objects into 10 classes and detects the objects in the test
sets. It contains natural images and helps implement the
image detection algorithms.
In this paper, Convolutional neural networks models
are implemented for image recognition on MNIST dataset
and object detection on the CIFAR-10 dataset. The
implementation of models is discussed and the
performance is evaluated in terms of accuracy. The model
is trained on an only CPU unit and real-time data
augmentation is used on the CIFAR-10 dataset. Along
with that, Dropout is used to reduce Overfitting on the
datasets.
The remaining sections of the paper are described as
follows: Section 2 describes a brief literature survey;
Section 3 describes the classifier models with details of
the techniques implemented. Section 4 evaluates the
performance of the model and describes the results.
Section 5 summaries the work with future works.
II. LITERATURE SURVEY
In recent years there have been great strides in
building classifiers for image detection and recognition on
various datasets using various machine learning
algorithms. Deep learning, in particular, has shown
improvement in accuracy on various datasets. Some of the
works have been described below:
Norhidayu binti Abdul Hamid et al. [3] evaluated the
performance on MNIST datasets using 3 different
classifiers: SVM (support vector machines), KNN (K-
nearest Neighbor) and CNN (convolutional neural
networks). The Multilayer perceptron didn't perform well
on that platform as it didn't reach the global minimum
rather remained stuck in the local optimal and couldn't
recognize digit 9 and 6 accurately. Other classifiers,
performed correctly and it was concluded that
performance on CNN can be improved by implementing
the model on Keras platform. Mahmoud M. Abu Gosh et
al. [5] implement DNN (Deep neural networks), DBF
(Deep Belief networks) and CNN (convolutional neural
networks) on MNIST dataset and perform a comparative
study. According to the work, DNN performed the best
with an accuracy of 98.08% and other had some error
rates as well as the difference in their execution time.
Youssouf Chherawala et al. [6] built a vote weighted
RNN (Recurrent Neural networks) model to determine the
significance of feature sets. The significance is
determined by weighted votes and their combination and
the model is an application of RNN. It extracts features
from the Alex word images and then uses it to recognize
handwriting. Alex krizhevsky [7] uses a 2-layer
Rahul Chauhan Kamal Kumar Ghanshala R.C Joshi
Graphic Era Hill University Graphic Era University Graphic Era University
Dehradun, India Dehradun, India Dehradun, India
chauhan14853@gmail.com kamalghanshala@gmail.com rcjoshi.geu@gmail.com
2018 First International Conference on Secure Cyber Computing and Communication (ICSCCC)
278978-1-5386-6373-8/18/$31.00 ©2018 IEEE

Convolutional Deep belief network on the CIFAR-10
dataset. The model built classified the CIFAR-10 dataset
with an accuracy of 78.90% on a GPU unit. Elaborative
differences in filters and their performance is described in
the paper which differs with every model.
Yehya Abouelnaga et al. [8] built an ensemble of
classifiers on KNN. They used KNN in combination with
CNN and reduce the Overfitting by PCA (Principal
Component Analysis). The combination of these two
classifiers improved the accuracy to about 0.7%.
Yann le Cunn et al. [1] give a detailed introduction to
deep learning and its algorithms. The algorithms like
Backpropagation with multilayer perceptron,
Convolutional neural networks, and Recurrent neural
networks are discussed in detail with examples. They
have also mentioned the scope of unsupervised learning in
future in Artificial intelligence.
Li Deng [10] details a survey on deep learning, its
applications, architectures, and algorithms. The
generative, discriminative and hybrid architectures are
discussed in detail along with the algorithms that fall
under the respective categories. CNN, RNN,
Autoencodes, DBN’s, RBM’s (Restricted Boltzmann
machines) are discussed with their various applications.
III. CLASSIFIER MODELS
A. Datasets
MNIST is the dataset used for image recognition i.e.
for recognition of handwritten digits [11]. The dataset has
70,000 images to train and test the model. The training
and test set distribution is 60,000 train images and 10,000
test images. The size of each image is 28x28 pixels (784
pixels) which are given as input to the system and has 10
output class labels from (0-9). Fig.1 shows a sample
picture from MNIST dataset [13].
CIFAR-10 is the dataset used for object detection
which is labeled a subset of 80 million tiny images [12].
The dataset has 60,000 32x32 pixel color images with 10
classes (airplane, automobile, bird, cat, deer, dog, frog,
horse, ship, truck ). Each class has 6000 images. The train
batch has 50,000 images and test batch has 10,000
images. The test batch for each class has 1000 images
which are randomly selected. Fig.2 shows sample pictures
from CIFAR-10 dataset [14].
B. CNN Models
Convolutional neural networks are deep learning
algorithms that take input images and convolves it with
filters or kernels to extract features. A NxN image is
convolved with a fXf filter and this convolution operation
learns the same feature on the entire image[18]. The
window slides after each operation and the features are
learnt by the feature maps. The feature maps capture the
local receptive field of th eimage and work with shared
Fig.1 A sample MNIST Dataset
Fig.2 A CIFAR-10 dataset images
weights and biases [15,21]. Equation (1) shows the size of
the output matrix with no padding and Equation (2) shows
the convolution operation. In order to preserve the size of
input image padding is used . In a ‘SAME’ padding the
output image size is same as input image size and a
“VALID” padding is no padding. The size of th eoutput
matrix with padding is depicted in equation (3).ܰܺܰ
ͳ ൅ ܨ െ ܰൌ ݂݂ܺכ (1)ܱ
ൌ ߪሺܾ ൅ σ σ ݓ௜ǡ௝
ଶ
௝ୀ଴
ଶ
݄௜ୀ଴ ௔ା௜ǡ௕ା௝ ሻ (2)ܰܺܰ
݂כ ݂כൌ ሺ ܰ൅ ʹ ܲെ݂ ሻȀሺ ݏ൅ ͳሻ (3)
Here, O is the output , P is the padding, s i sthe stride,
b is the bias, σ is the sigmoidal activation function, w is a
3x3 weight matrix of shared weights and݄ ௫ǡ௬ is the input
activation at position x, y. CNN has application in fields
of large scale image recognition [17], Emotion detection
through speech [9] [19] facial expression recognition [20],
biometric systems, genomics and many others.
C. CNN model for MNIST dataset
The CNN model for MNIST dataset is shown in figure
(c). The input image is a vector with 784 pixel values. It is
input into the convolutional model where the convolution
layers along with filters generate th efeature maps using
the local receptive field. The pooling and fully connected
layers follow the convolution layers. Dropout is
2018 First International Conference on Secure Cyber Computing and Communication (ICSCCC)
279

introduced after each convolutional layer. The pooling
layers simplify the output after convolution. There are two
types of pooling: Max pooling and L2 pooling. In max
pooling the maximum activation output is pooled into a 2
x 2 input region and L2 pooling takes the square root of
the sum of squares of the activation 2x2 region. Finally,
the fully connected layers connect each layer of max
pooling layer to the output neurons. The architecture of
the developed model is as follows:
 Batch Size (training): 128, Batch size (test): 256,
Number of epochs:10
 Dropout: Yes
 Optimizer: RMS prop, Learning rate=0.001, the
parameter β : 0.9
 Keep_prob:0.8
The architecture of the model is :
Convolution_layer 1 → Relu → Max_pool → dropout →
Convolution_layer 2 → Relu → Max_pool → dropout →
Convolution_layer 3 → Relu → Max_pool →
fully_connected → dropout → output_layer → Result
The input is a 28X28 image which passed to filters to
generate fetaure maps. . The first filter is of size 5x5x1x32
(32 features to learn in the first hidden layer), 3x3x32x64
for the second convolution layer (64 features to learn from
second hidden layer), 3x3x64x128 for the third layer,
(128*4*4,625) for the fourth layer and (625,10) for the
last layer. The stride is 1 for convolution layer and 2 for
max-pooling layers. Stride defines the number of block to
move forward after one calculation. Generally, the value
of stride for convolution layer is 1 and for pooling layer is
2. The accuracy of this model is 99.6%.
Fig. 3. CNN model for MNIST dataset
D. CNN model for CIFAR-10 dataset
The CNN model for CIFAR-10 dataset is as follows:
 Batch size: 32, Number of epochs:50
 Optimizer: RMS Prop, Decay rate= 1e-6
 Data augmentation: True, Rotation ; in range of
0-180, horizontal flip: TRUE, Vertical flip:
FALSE
 Dropout : True
The architecture is as follows:
Conv1→Relu→Conv2→Relu→Max_pooling→Dropout
→Conv3→Relu→Conv4→Relu→Max_pooling→Dropo
ut→Conv5→Relu→Max_pooling→Dropout→Flatten→
Dense→Relu→Dropout→Dense→Softmax→Result
Fig.4 shows the architecture of the model. The 32X32
input image is given to the model where the first
convolutional layer learns 32 features through a 5x5 filter
and ‘same’ padding. After the activation, another
convolutional layer is stacked up that learns 64 features
through a 5x5 filter. Then Relu activation is added and
forwarded to max pooling layer. After the max pooling
layer a dropout is implemented with the dropout of 0.25.
This entire layer is repeated again with 3x3 filters and the
conv 3 layer this time learns 64 features similar to conv 4
features. All the remaining parameters are same. The conv
5 layer learns 64 features with a 3x3 filter followed by
Relu, max-pooling and dropout. Then the output is
flattened and we have a fully connected layer with 512
outputs. The dropout is again applied as 0.5 and denser to
number of output classes i.e. 10 and passed to the softmax
layer for the final output. The accuracy of the model is
80.01 % on a CPU unit on test dataset.
.
Fig.4 CNN Model for CIFAR-10
E. Relu non-linearity activation function
There is a wide range of activation function available
when training neural network models. The mainly used
activations are sigmoid, tanh, relu and leaky relu. The relu
non linearity is a popular activation function used in deep
learning algorithm sand has replaced the use of sigmoidal
activation function which is generally used for binary
classification techniques. Relu non linearity ˆሺšሻ ൌ
ƒšሺͲǡ šሻ works several times faster than the tanh non –
linearity ሺ‡୸ െ ‡ି ୸ ሻȀሺ‡୸ ൅ ‡ି ୸ ሻ , and results the output of
the activation as either 0 or a positive number. With relu it
is easier to train larger neural networks. Fig.5 shows the
graph of relu non-linearity [16].
F. Overfitting and under fitting
Another aspect of training a deep neural network is the
issue of high bias (resulting in underfitting) or high
variance (resulting in overfitiing). When the data is
Input
28*28
Image
Conv1
28*28*32
with 5*5,
s=1 filter
Max_p
ool
(2*2,
s=2)
Max_poo
l
(2x2,
s=2)
Conv2
with 3x3,
s=1 filter
Reshaped
FC (625,
10)
Conv3
with 3x3,
s=1 filter
Fully_conn
ected
Reshape to
(128*4*4,6
25)
Output Layer
Conv 1(32,5x5 filter,
stride=1) →Relu→Conv2
(64,5x5 stride=1)
→Relu→Max-Pool
(stride=2) →dropout (0.25)
Conv3 (64,3X3,
stride=1)
→relu→conv4
(64,3x3, stride=1)
→relu→dropout
(0.25)
A
A
Conv 5(64, 3x3,
stride=1) →relu
Max_pool
(stride=2, 2x2)
→dropout (0.25)
Flatten→dense
(512)
→relu→dropou
t (0.5)
→Dense→soft
max
Result
2018 First International Conference on Secure Cyber Computing and Communication (ICSCCC)
280

underfitting is not generalizing or fitting the data points
well. In case of high bias, which is on training batch the
network needs to be trained longer and have much bigger
network of hidden layers. In case of overfitting, the data
has high variance i.e. it is generalizing too well on the test
sets. To reduce high variance, regularization techniques
and data augmentation techniques can be implemented.
Fig.5. Relu non-linearity
G. Dropout to reduce overfitting
Dropout is a regularization technique which is used to
reduce overfitting [7]. In dropout, the network deads some
of its nodes randomly based on a parameter. The
probability parameter determines whether the node should
remain in the network or not. Keep_prob is the probability
parameter to keep the hidden nodes in the network. The
activation is unaffected during this process as it only
determines whether to keep the node in the network or
not.
H. Data Augmentation
Another technique to reduce overfitiing is to train the
data on large datasets. If dataset is limited the dataset can
be artificially created by data augmentation techniques
[7]. The data augmentation techniques include distortion
and altering of data images for processing to get more
data. Some of the techniques are:
 Mirroring – The images are flipped and laterally
inverted.
 Random cropping- Cropping some parts of the
image and creating subsets from the main image.
 Rotation- This includes rotating the images in
any direction at various angles and generating
new images.
 Color shifting- Shifting the RGB pixel values of
the image to get a new coloured image.
I. RMS prop optimizer and learning rate
RMS prop or root mean square prop is an optimizer
which works on the root mean square value of the change
in gradients. The change in weights and bias determine
the gradient parameters with help of rms value. The
learning rate determines the steps the algorithm will take
to converge to the global minimum. If the learning rate is
too high the algorithm faces exploding gradient problem
i.e. it takes larger steps and fails to converge at the local
minimum. If the learning rate is too small it faces a
vanishing gradient problem i.e. the gradient takes smaller
steps and ultimately becomes so small that the changes in
weights are insignificant. Thus the learning rate is a hyper
parameter that needs to be finely tuned and can be done
with the help of learning rate decay. In learning rate
decay, the learning rate decays exponentially after every
epoch.
IV. RESULT ANALYSIS
The results of the experiments are as shown below:
 CNN Model for MNIST dataset: Accuracy
99.6% shown in figure 6
Fig.6 Accuracy of CNN model on MNIST in 10 epochs
The result shows that as the number of epochs get increased
and the best accuracy achieve in recognizing the digits on
MNIST data set is 99.6 % with 10 epochs.
Fig. 7 Accuracy of the CNN model in 50 epochs
CNN model for CIFAR-10 dataset: Accuracy of 80.17%
on test set as shown in figure 7.
V. CONCLUSION
The article discusses various aspects of deep learning,
CNN in particular and performs image recognition and
detection on MNIST and CIFAR -10 datasets using CPU
unit only. The accuracy of MNIST is good but the
accuracy of CIFAR-10 can be improved by training with
larger epochs and on a GPU unit. The calculated accuracy

on MNIST is 99.6% and on CIFAR-10 is 80.17%. The training accuracy on CIFAR-10 is 76.57 percent after 50

epochs. The accuracy on training set may also be
improved further by adding more hidden layers. And this
system can be implemented as a assistance system for
machine vision for detecting nature language sysmbols.
REFERENCES
[1] Yann LeCun, Yoshua Bengio, Geoffery Hinton, “Deep Learning”,
Nature, Volume 521, pp. 436-444, Macmillan Publishers, May
2015.
[2] Norhidayu binti Abdul Hamid, Nilam Nur Binti Amir Sjarif,
“Handwritten Recognition Using SVM, KNN and Neural
Network”, www.arxiv.org/ftp/arxiv/papers/1702/1702.00723
[3] Cheng-Lin Liu
כ, Kazuki Nakashima, Hiroshi Sako, Hiromichi
Fujisawa, “Handwritten digit recognition: benchmarking of state-
of-the-art techniques”, ELSEVIER, Pattern Recognition 36 (2003)
2271 – 2285).
[4] Ping kuang, Wei-na cao and Qiao wu, “Preview on Structures and
Algorithms of Deep Learning”, 11th International Computer
Conference on Wavelet Actiev Media Technology and
Information Processing (ICCWAMTIP), IEEE, 2014.
[5] Mahmoud M. Abu Ghosh ; Ashraf Y. Maghari, “A Comparative
Study on Handwriting Digit Recognition Using Neural Networks”,
IEEE, 2017.
[6] Youssouf Chherawala, Partha Pratim Roy and Mohamed Cheriet,
“Feature Set Evaluation for Offline Handwriting Recognition
Systems: Application to the Recurrent Neural Network,” IEEE
Transactions on Cybernetics, VOL. 46, NO. 12, DECEMBER
2016.
[7] Alex Krizhevsky, “Convolutional Deep belief Networks on
CIFAR-10”. Available: https://www.cs.toronto.edu/~kriz/conv-
cifar10-aug2010.pdf.
[8] Yehya Abouelnaga , Ola S. Ali , Hager Rady , Mohamed
Moustafa, “ CIFAR-10: KNN-based ensemble of classifiers”,
IEEE, March 2017.
[9] Caifeng Shan, Shaogang Gong, Peter W. McOwan, “Facial
expression recogniton based on Local binary patterns: A
comprehensive study”, ELSEVIER, Image and Vision Computing
27, pp. 803-816, 2009.
[10] Li Deng, “A tutorial survey of architectures, algorithms, and
applications of Deep Learning”, APSIPA Transactions on Signal
and Information Processing (SIP), Volume 3, 2014.
[11] Yann LeCun, Corinna Cortes and Christopher J.C. Burges, “The
MNIST Database of handwritten digits”. Available:
http://yann.lecun.com/exdb/mnist/ - MNIST database
[12] The CIFAR-10 dataset. Available:
https://www.cs.toronto.edu/~kriz/cifar.html
[13] MNIST dataset introduction, 2017. Available:
http://corochann.com/mnist-dataset-introduction-1138.html
[14] Robust Vision Benchmark. Available:
https://robust.vision/benchmark/about/
[15] Neural Network and Deep Learning. Available:
http://neuralnetworksanddeeplearning.com/chap6.html
[16] Convolutional Neural Networks for Visual Recoginition. Available:
http://cs231n.github.io/neural-networks-1/
[17] Krizhevsky, Sutskever and Hinton, “ImageNet classification with
deep convolutional neural networks”, Advances in Neural
Information Processing Systems 25 (NIPS 2012), pp. 1106–1114,
2012.
[18] Zeiler, M. D. and Fergus, “Visualizing and understanding
convolutional networks”. European Conference on Computer
Vision, vol 8689. Springer, Cham, pp. 818-833, 2014.
[19] Zhengwei Huang, Min Dong, qirong Mao and Yongzhao Zhan,
“Speech Recognition using CNN”, IEEE/ACM Transactions on
Audio, Speech and Language Processing, pp. 1533-1545, Volume
22, Issue 10, 2014, http://dx.doi.org/10.1145/2647868.2654984.
[20] Shima Alizadeh and Azarr Fazel, “Convolutional Neural networks
for Facial Expression recognition”, Computer Vision and Pattern
Recognition, Cornell University Library, ArXiv:1704.06756v1, 22
April, 2017, arXiv.org.1704.06756.pdf.
[21] Christian szegedy, Wei liu, Yangqing Jia et al., “Going Deeper with
Convolutions”,Conference on Computer Vision and Pattern
Recognition (CPVR) , IEEE explorer, Boston, MA, USA, 2015.