Page 40 - Kỷ yếu hội thảo khoa học lần thứ 12 - Công nghệ thông tin và Ứng dụng trong các lĩnh vực (CITA 2023)

24

Fig. 1. Automatic detection and classification of sounds for warning system

2 Background

2.1 Deep learning

Deep learning is a part of Machine Learning where software learns and improves itself
through algorithms. Deep Learning is built on a lot more complex concepts compare
to Machine learning, mainly working with artificial neural networks to capture the
thinking and abilities like human.

Supervised learning is the most common form of Machine learning. An algorithm
which the model must be trained with the defined data before implementation. The
process is supervised by performed with labeled training data that human gives. One of
a main type of supervised learning is Classification.

Classification
often given a discrete result (yes or no, positive, or negative). For example, to predict
an input sound is a gunshot or not, the model define that sound will be positive with
gunshot label or negative one and repeat that process with another label until it finds
out the positive label (True label) for that sound based on related signal. The algorithm
determines how similar (normal) or dissimilar (abnormal) of the input sound compared
to signals learned from the training process.

2.2 Convolutional Neural Network

Convolutional Neural Network is a type of Deep learning, a neural network that its
foundation of many convolutions layer (Fig. 2). These layers are connected through
individual connections between nodes, each containing a weight: the value that
represents the strength between the connection of two nodes.

CITA 2023 ISBN: 978-604-80-8083-9