Abstract:
Masks used to avoid the spread of the infection to be used especially in closed
and crowded environments, the people should continue wearing face mask which can
protect from seasonal allergies. It reduces the symptoms associated with seasonal allergies.
A dataset consisting of images of people with masks or without mask is created and used
in this study. The images were collected in total 7553 from Kaggle and own since a dataset
that relates to faces has many privacy issues attached it. We converted image size into 128
*128 and transformed into RGB. The study considers an approach that aggregates
Convolution Neural Networks (CNN) deep learning techniques and classical ML
classifiers. To classify images into masked and unmasked own architecture on CNN was
used to extract unique features, then Support Vector Machine (SVM) used to classify the
image. Purpose of the hybrid modelling on CNN, it recognizes these advanced features
from the sample data, thus reducing the workload of developing a new feature extractor
for such problem and when the number of data is small, CNN do not work well. In order
to sort out that issue we bring SVM algorithm also here which can be possessing high
accuracy even with less data. The problem for our proposed model is to learn the
interpretation of various features in images and classify accordingly. By analysing various
architectures on CNN, own model created in leveraging the spatial information in images.
With this concept, the own architecture designed because some of existing architectures
are complex, some pertinent information may be lost there during feature extraction phase
and the computation time is considerable. The pervious Custom CNN architecture Model
training accuracy reached 94% and Validation accuracy 96%. When the CNN was used as
a feature extractor, the SVM classifier was demonstrated to be the best combining
counterpart, providing the best synergy effect in terms of accuracy. This indicated that the
proposed fusion achieved superior recognition accuracy of 97.35 % compared to the
previous approach. We used 1511 data for testing from that 10% used for validation.
