Mirror Mosaicking: A Novel Approach to Achieve High-performance

Classification of Gases Leveraging Convolutional Neural Network

S. N. Chaudhri

and N. S. Rajput

Indian Institute of Technology (BHU), Varanasi-221005, UP, India

Keywords: Mirror Mosaicking, Gas Sensor Array, Electronic Nose, Gas Classification, Pattern Recognition,

Convolutional Neural Network.

Abstract: Limited dimensionality of the dataset obtained from an electronic nose (EN) is due to the number of elements

in the sensor array used generally in the range of 4-8 elements only. Further, large number of sensor data can

be generated by sampling the sensor responses both during the transient and steady states. The lower-

dimensionality of sensor data prohibits the use of a convolutional neural network (CNN)-based pattern

recognition techniques because the kernels of a CNN cannot be used on the obtained sample vectors to extract

the features. In this paper, we have proposed a novel approach to enhance the data dimensionality keeping the

sensor response characteristics absolutely unaltered. By leveraging the concept of mirror mosaicking

technique, we have upscaled the input sample vectors into a 6×6 2-D input arrays to train the shallow CNN.

Using the proposed approach, all the 16-unknown steady-state test samples classified accurately which are

not used during the training. Moreover, the parameters of the classification report viz., Precision, Recall, and

F1 score also obtained with a fraction value of 1.00. The proposed technique is a generic approach that can

be used to classify various low-dimensional datasets obtained from various sensor arrays in various fields.

1 INTRODUCTION

In the current scenario, Artificial Intelligence (AI),

Machine Learning (ML), Deep Learning (DL), and

modern Pattern Recognition (PR) techniques are

finding their applications in almost all the research

areas; for delivering better results. An electronic nose

(EN) is the mimicry of the olfactory system that is a

popular topic of research as a multidisciplinary area.

The word multidisciplinary represents the wide area

of applications of EN related to different industries.

Various traditional pattern recognition approaches

have been used for the classification of gases or

odors, as described by (Santos et al., 2017; Fujinaka

et al., 2008; Hodgins and Simmonds, 1995; Tang et

al., 2010; Keller et al., 1995; Rodrguez et al., 2010;

Capelli et al., 2014; Kızıl et al., 2017; Chen et al.,

2013). The EN is a system that contains a gas sensor

array consisting of few sensors typically 4 to 16.

Moreover, data pre-processing and pattern

recognition modules are the main parts of any EN

system (Arshak et al., 2004). The EN system can be

https://orcid.org/0000-0002-5436-2977

https://orcid.org/0000-0002-1650-011X

made more selective for analytes under observation,

using an array of sensors (Zhang et al., 2017). A gas

sensor array logically has more than one sensor

element to enhance the selectivity of the system. If

there are fewer numbers of sensors in a gas sensor

array, the resulting response dataset has a feature

vector of limited size for each sample. A concept of

mirror mosaicking technique is proposed in this work

to broaden the applicability of deep learning pattern

recognition techniques for automatic feature

extraction and classification of small gas sensor array

responses. Subsequently, any gas sensor array

response can be analyzed using the convolutional

neural network (CNN) at the sample level irrespective

of the size of the gas sensor array. The feature vector

of any sample is obtained from the respective gas

sensor array response having the length equal to the

number of the sensor elements. Each pattern

recognition technique requires a specific input format

or length of the feature vector. For example, various

dimensional versions viz., 1-D, 2-D, and 3-D based

on the type of operation of the convolution of CNN

Chaudhri, S. and Rajput, N.

Mirror Mosaicking: A Novel Approach to Achieve High-performance Classiﬁcation of Gases Leveraging Convolutional Neural Network.

DOI: 10.5220/0010251500860091

In Proceedings of the 10th International Conference on Sensor Networks (SENSORNETS 2021), pages 86-91

ISBN: 978-989-758-489-3

require a 1-D vector, 2-D array, and 3-D array (Eren,

2017; Yamashita et al., 2018). Therefore, each

version has its limitation of input sample

representation. A 2-D CNN has vast popularity in the

area of computer vision and image processing.

However, any of the versions of CNN can be used

depending on the available dimension of the sample.

In this work, data pre-processing and classification

parts of an EN system has been demonstrated. The

pre-processing has been used to obtain the mosaicked

sample and later on the classification task is

implemented using a convolutional neural network.

The rest content has been organized in the

following sections. The proposed mirror mosaicking

technique has been explained in Section 2. While

Section 3 describes the designed convolutional neural

network to generalize the proposed approach.

Moreover, Section 4 and Section 5 are dedicated to

the used material and obtained results, respectively.

Lastly, the conclusions have been summed up in

Section 6.

2 MIRROR MOSAICKING

2.1 Need of Mirror Mosaicking

As discussed in the previous section, the particular

dimensional version of CNN requires a specific

format of the input. The 1-D CNN requires input as a

feature vector. The 1-D CNN significantly can be

applied if the feature vector has sufficient length. The

length of the feature vector proportionally depends on

the number of sensor elements in the gas sensor array.

As quoted earlier, a gas sensor array has more than

one sensor element, then the least possible array must

have two gas sensors elements. Let us suppose we

have the smallest gas sensor array, then the feature

vector of this array will have a length of 2 units. This

length of the feature vector is insufficient to explore

the significance of the 1-D CNN. Similarly, the 2-D

CNN needs input in the form of a 2-D/3-D matrix

(grayscale/color image). A feature vector of the

smallest sensor array of length 2 units is incompatible

for converting into a 2-D matrix. Therefore, the

sample obtained from this sensor array cannot be feed

as input into the 1-D CNN or the 2-D CNN and the

subsequent higher dimensional version 3-D CNN.

Consequently, a sample obtained from the least

feasible gas sensor array or the gas sensor array with

two sensor elements is insufficient to deal with the 1-

D CNN and the 2-D CNN. Moreover, these samples

will also be insufficient to the subsequent higher

dimensional version of CNN that is 3-D CNN. Hence

a technique called mirror mosaicking has been

proposed so that each sample obtained from any

sensor array can be analyzed using 2-D CNN. The

popularity of 2-D CNN has been proved worldwide

by computer vision and image classification

applications.

2.2 Implementation of Mirror

Mosaicking

The proposed mirror mosaicking is the approach in

which each sample vector obtained from any gas

sensor array can be converted into a 2-D matrix of

significant size so that a 2-D CNN can be applied at

the sample level (Chaudhri et al., 2020). First of all,

the original sample feature vector is converted into

the square 2-D matrix using zero-padding if required.

Subsequently, the mirror mosaicking technique is

applied to this square 2-D matrix to obtain the desired

mosaicked sample compatible with the 2-D CNN.

Figure 1: Directions of Mirrors.

Let's assumed that eight mirrors are placed around

the square 2-D matrix obtained from the original

sample feature vector using zero-padding if required.

The corresponding directions of the mirrors are

shown in Figure 1. The obtained corresponding

mirror images of the square 2-D matrix at the center

are mosaicked with it at their respective locations to

obtain the desired mosaicked sample. Thus, the

obtained mosaicked sample of the corresponding

original sample now compatible as input to the 2-D

CNN. Further, the depth of the used 2-D CNN is the

thing to note down. There is no restriction for the

depth of the network based on the fully connected

layers provided the increasing number of layers

makes the model complex. But concerning the

convolutional and the pooling layers, the network

depth depends on the size of the input. Considering

the case of the smallest feasible gas sensor array a

shallow convolutional neural network (SCNN) has

Mirror Mosaicking: A Novel Approach to Achieve High-performance Classiﬁcation of Gases Leveraging Convolutional Neural Network

been designed that can be extended according to the

size of the input by adding more convolutional or

pooling layers. The word shallow indicates that few

convolutional layers in the designed 2-D CNN are

two according to the size of the input.

3 SHALLOW CONVOLUTIONAL

NEURAL NETWORK (SCNN)

As quoted in Section 2, the designed shallow

convolutional neural network (SCNN) has two

convolutional layers. Depending on the size of the

available input sample, convolutional and pooling

layers can be increased in the designed model. Mainly

pooling layers are used in those deep networks which

are specifically designed to deal with image-related

tasks because images are made of large 2-D matrices.

In our work, SCNN is designed to classify the gases

using gas sensor array responses which have a limited

sample size. But the SCNN can be extended to

classify the responses obtained from the gas sensor

array having any number of sensor elements. The

used SCNN has the following layers:

• Convolutional Layers (Input Layer)

• Flatten Layer

• Dense or Fully Connected Layers

• Dropout Layers

• Softmax Layer (Output Layer)

The SCNN with the mentioned layers has been

designed considering the smallest gas sensor array. A

brief theoretical introduction for the basic layers is

given below (Bhandare et al., 2016):

Convolutional Layers.

The model attains the leading significance of

automatic feature extraction by this layer. In this

layer, the number of kernels is initialized, which are

used to produce the same number of feature maps.

The feature maps are obtained from the convolution

of the input and the kernel. All the stacked feature

maps are forwarded to the next layer in the form of

input.

Pooling Layers.

It is used for down-sampling. There are three basic

types of pooling namely max pooling, min pooling,

and average pooling. Out of the three types, max

pooling is used widely.

Flatten and Fully Connected Layers.

The flatten layer is used after all the used

convolutional and pooling layers to convert the output

feature map into the vector format. Subsequently, the

dense layers or the fully connected layers are used. A

fully connected layer is that in which each neuron is

connected to each neuron of the previous and next

layer.

Softmax Layer.

The softmax layer is the output layer in which the

number of neurons must equal to the number of

targets. In this layer, for each input, the membership

fraction corresponding to each target is obtained.

Dropout and Normalization Layers.

The dropout and the normalization, both layers are

used to get rid of overfitting. In the dropout layer, a

dropout amount is initialized to discard the neurons

containing the value less than or equal to the dropout

amount. The value of the dropout amount always lies

between 0 and 1. While, the normalization is required

for improving the speed, performance, and stability of

the network, in the complex network models.

After the brief introduction of the layers, the

architecture of the designed SCNN can be explained

easily. This model contains two convolutional layers,

the flatten layer, a fully connected layer, and the

softmax layer. The model termed as shallow network

instead of a deep network, due to the use of only two

convolutional layers. A schematic diagram of the

proposed network is shown in Figure 2. Since the

input size is limited, so there is no need for down-

sampling. Accordingly, the pooling layer has not been

used in the proposed network. If the CNNs are used

to deal with image data (large matrices), then pooling

layers are essentially used in the designed network.

4 USED MATERIAL

In this work, the material used to verify the proposed

methodology has been taken from a thick film gas

sensor array. This gas sensor array consists of the

following four sensors:

• Cadmium sulfide (CdS)

• Molybdenum Oxide (MoO)

• Tin Oxide (SnO

)

• Zinc Oxide (ZnO)

Four gases had been exposed to this array in the

ambiance of Nitrogen (N2). These gases are

mentioned below:

• Acetone (CH

COCH

)/ACE

• Carbon Tetrachloride (CCl

)/CAR

• Ethyl Methyl Ketone (C

COCH

)/EMK

• Xylene (CH

)/XYL

SENSORNETS 2021 - 10th International Conference on Sensor Networks

The steady-state response of the mentioned gas sensor

array has been taken to verify the results. The SCNN

has been trained with exclusive samples from the test

samples. Later on, the classification performance is

obtained using test samples. The performance of the

classifier model on a smaller dataset signifies the

applicability of the proposed methodology using the

mirror mosaicking technique. Moreover, our

methodology is not limited to the number of samples

or number of the sensor elements in the gas sensor

array. It can be used for any gas sensor array. The

explained details about the gas sensor array and its

response are given in (Nayak et al., 1994; Rajput et

al., 2010).

Figure 2: Schematic block diagram of SCNN.

5 RESULTS AND DISCUSSION

The results obtained from the classifying network

SCNN after applying on the steady-state response

dataset has been discussed in this section. The

parameters that are used to tune the classifying

network have been listed in Table 1. The

classification of all the instances proves the

applicability and efficiency of our proposed

technique. The classification report is shown in Table

2. Precision, Recall, and F1-score metrics are shown

in the report. The expressions for all the metrics have

been given in equation (1), (2), and (3) respectively.

All the aforesaid metrics have been calculated based

on the confusion matrix. A confusion matrix is a

square matrix of size equal to the targets. The

elements of this matrix show the description of the

reference points and the corresponding predicted

outputs.

Precision 



(1)

Recall 



(2)

F1 Score 

2 ∗ Precision ∗ Recall

Precision  Recall

(3)

Accurac



TP  TN



TP  TN



FPFN

(4)

The overall accuracy has been given by equation (4).

The term TP, TN, FP, FN in equation (1), (2), (3), and

(4) represent the correctly predicted positive values,

correctly predicted negative values, actual negative

predicted positive, and actual positive predicted

negative values respectively. Using equation (4) the

overall accuracy has been obtained equal to 1.00. In

terms of the percentage, it is obtained as 100%, as

shown in Table 2. Moreover, the overall classification

accuracies for the used dataset have been given in

Table 3 using various classifying techniques.

The proposed technique is a comprehensive

approach that can also be used to classify the dataset

obtained from the transient response of a gas sensor

array. There are two ways to deal with the transient

response. Firstly, the last observation of transient

response can be considered as a steady-state response.

Secondly, the averaged transient response can be

considered as a steady-state response. In this way, the

huge computational cost can be reduced up to a very

low cost. Moreover, the pre-processing and

classification procedure will remain the same.

Table 1: Model Parameters.

Size of input samples 6×6

Convolutional layer 1

Number of kernels

Size of kernels Activation

3×3

tanh

Convolutional layer 2

Number of kernels

Size of kernels

Activation

3×3

tanh

Flatten layer ( )

Fully connected layer 1

Number of neurons Activation

tanh

Dropout layer 1 0.25

Softmax layer

Number of targets

Activation

softmax

Optimizer

Learning rate

Adam

0.001

6 CONCLUSIONS

The proposed technique provides all the samples

well-classified that proves the significance of the

mirror mosaicking technique. The proposed

technique can be used to classify any gas sensor array

response using well-known CNN-based pattern

recognition techniques. In a nutshell, the points of

significance of this paper can be stated as follows:

Mirror Mosaicking: A Novel Approach to Achieve High-performance Classiﬁcation of Gases Leveraging Convolutional Neural Network

• A new technique "Mirror Mosaicking" of data

pre-processing has been proposed.

• The dataset obtained from the response of the least

feasible size gas sensor array can be classified

using a convolutional neural network using mirror

mosaicking.

Table 2: Classification Accuracies for Classical Machine

Learning Datasets using Convolutional Neural Network

based on Mirror Mosaicking Approach.

Datasets

Train/Test

Samples

Overall Test

Accuracy (%)

IRIS Dataset 120/30 100

Wine Dataset 112/66 98.48

Parkinson’s

Dataset

136/59 100

Table 3: Classification Report.

Precision Recall

Score

Support

ACE 1.00 1.00 1.00 2

CAR 1.00 1.00 1.00 3

EMK 1.00 1.00 1.00 6

XYL 1.00 1.00 1.00 5

Avg./

Total

1.00 1.00 1.00 16

Test

Accuracy

100%

The proposed technique is a generic approach that

can be used to classify any other non-imaging

datasets, obtained from any other sensor arrays in

various fields. For example, various classical

machine learning datasets viz., iris data, wine data,

Parkinson's disease data (Dua et al., 2019; Little et al.,

2007), etc. can be classified accurately by using the

proposed technique. The classification accuracies for

these datasets have been given in Table 4 which have

been obtained using convolutional neural networks

based on the mirror mosaicking approach.

Table 4: Classification Accuracies using Various

Classifiers.

Classifier Overall Accuracy (%)

KNN 87.50

Linear SVM 81.25

RBF SVM 87.50

Random Forest 93.75

Naïve Bayes 87.50

ACKNOWLEDGEMENTS

We acknowledge the administrative, technical and

financial support received in parts from NCC LAB,

Department of Electronics Engineering, IIT (BHU),

INDIA (Grant No. IS/ ST/ EC-13-14/02) and from

M/s IBM, INDIA (Grant No. R&D/ IBM/

SBApp/Electronics/ 15-16/ 07).

REFERENCES

Arshak, K., Moore, E., Lyons, G. M., Harris, J., and

Clifford, S. (2004). A review of gas sensors employed

in electronic nose applications. Sensors Rev., vol. 24,

no. 2, pp. 181-198.

Bhandare, A., Bhide, M., Gokhale, P., and Chandavarkar,

R. (2016). Applications of convolutional neural

networks. International Journal of Computer Science

and Information Technologies, pp. 2206-2215.

Capelli, L., Sironi, S., and Rosso, R. D. (2014). Electronic

noses for environmental monitoring applications.

Sensors, 14 (11), pp. 19979-20007.

Chaudhri, S. N., Rajput, N. S., and Singh, K. P. (2020). The

novel camouflaged false color composites for the

vegetation verified by novel sample level mirror

mosaicking based convolutional neural network.

InGARSS 2020, to be published (Accepted).

Chen, S., Wang, Y., and Choi, S. (2013). Applications and

technology of electronic nose for clinical

diagnosis. Open Journal of Applied Biosensor 2, pp.

39-50.

Dua, D., and Graff, C. (2019). UCI Machine Learning

Repository [http://archive.ics.uci.edu/ml]. Irvine, CA:

University of California, School of Information and

Computer Science.

Eren, L. (2017). Bearing fault detection by one-dimensional

convolutional neural networks. Mathematical Problems

in Engineering, 1–9.

Fujinaka, T., Yoshioka, M., Omatu, S., and Kosaka, T.

(2008). Intelligent electronic nose systems for fire

detection systems based on neural networks. The 2nd

International Conference on Advanced Engineering

Computing and Applications in Sciences ADVCOMP,

pp. 551-554.

Hodgins, D., and Simmonds, D. (1995). The electronic nose

and its application to the manufacture of food products.

J. Automatic Chemistry, vol. 17, no. 5, pp. 179-185.

Keller, P. E., Kangas, L. J., Liden, L. H., Hashem, S., and

Kouzes, R. T. (1995). Electronic noses and their

applications. IEEE Northcon/Technical Applications

Conference (TAC’95), pp. 116-119.

Kızıl, U., Genç, L., and Aksu, S. (2017). Air quality

mapping using an e-nose system in Northwestern

Turkey. Agronomy Research.15(1): 205-218.

Little, M. A., McSharry, P. E., Roberts, S. J., Costello, D.,

and Moroz, I. M. (2007). Exploiting nonlinear

SENSORNETS 2021 - 10th International Conference on Sensor Networks

recurrence and fractal scaling properties for voice

disorder detection. Biomed. Engg., vol. 6, no. 23.

Nayak, M. S., Dwivedi, R., and Srivastava, S. K. (1994).

Sensitivity and response times of doped tin oxide

integrated gas sensors. Microelectron. J. 25, 17–25.

Rajput, N. S., Das, R. R., Mishra, V. N., Singh, K. P., and

Dwivedi, R. (2010). A neural net implementation of

SPCA pre-processor for gas/odor classification using

the responses of thick film gas sensor array, Sensors

and Actuators B: Chemical, vol. 148, pp. 550–558.

Rodrguez, J., Durn, C., and Reyes, A. (2010). Electronic

nose for quality control of Colombian coffee through

the detection of defects in cup tests. Sensors, vol. 10,

no. 1, pp. 36-46.

Santos, J. P., Lozano, J., and Aleixandre, M. (2017).

Electronic noses applications in beer technology. Brew.

Technol. 177.

Tang, K.-T., Chiu, S.-W., Pan, C.-H., Hsieh, H.-Y., Liang,

Y.-S., and Liu, S.-C. (2010). Development of a portable

electronic nose system for the detection and

classification of fruity odors", Sensors, vol. 10, pp.

9179-9193.

Yamashita, R., Nishio, M., Do, R. K. G., and Togashi, K.

(2018). Convolutional neural networks: an overview

and application in radiology. Insights Imaging 2018.

Zhang, Y., Zhao, J., Du, T., Zhu, Z., Zhang, J., and Liu, Q.

(2017). A gas sensor array for the simultaneous

detection of multiple VOCs. Sci. Rep., 7, p. 1960.

Mirror Mosaicking: A Novel Approach to Achieve High-performance Classiﬁcation of Gases Leveraging Convolutional Neural Network