Page 198 - Kỷ yếu hội thảo khoa học lần thứ 12 - Công nghệ thông tin và Ứng dụng trong các lĩnh vực (CITA 2023)

182

monitoring areas, and the service life of the sensors is not durable due to environmental
influences and geographical conditions. For warning systems using satellites can
cover a wide area, but there are some limitations such as the low resolution of images,
high cost, and influenced of weather. In fact, a highly regarded method of forest fire
detection is to rely on images and videos collected from CCTV cameras. Therefore,
image processing techniques are studied and applied a lot in detecting fire fires.
Recently, algorithms using deep learning have been of great interest because they
can detect fires quickly with high accuracy. Image recognition algorithms based on
convolutional neural networks (CNN) can efficiently learn and extract complex image
features. As a result, some researchers have applied CNN to fire detection through
imaging. In [1], the team detected fires based on a GoogleNet model with data
collected from surveillance cameras. In the paper [2], [3], [4], the researchers applied
the CNN network in smoke detection. In [5], the team proposes to improve the CNN
network to detect fires in real-time. The detection of fire and escape with little data
was studied and proposed by A. Namozov's research team, which corrected
overfitting using data augmentation techniques and generative adversarial networks
[6]. In 2018, research teams compared AlexNet, VGG, Inception, ResNet, etc. models
and developed smoke and flame detection algorithms [7, 9]. In 2017, Muhammad's
team detected early fires using the CNN network with self-generated datasets,
fine-tuned data collection cameras, and a proposed channel selection algorithm for
cameras to ensure data reliability [8]. In [10] the authors studied the feasibility of
network models with data collected through unmanned aerial vehicles (UAVs).
Y.Luo's team [1 1] came up with a moving object detection method to create
suggested zones based on background motion updates, then used the CNN algorithm
to detect smoke in these areas. In [12], the authors used Gaussian (MOG) to
distinguish the background and foreground, used a floor model to identify proposed
smoke regions, and then applied the CaffeNet network to detect smoke. The method
of using the suggested color point of the fire area, the AlexNet network to detect fire
is given in the paper [13]. In 2020, the authors detected Forest fire using an LBP color
signature combined with a deep learning network to detect smoke and fire with
datasets collected from overhead cameras [14]. The authors in [15] offer a combined
solution using LSTM and YOLO models to detect smoke in forest fire environments.
The LSTM model reduces the number of layers and obtains better results in smoke
detection. Another method in [16] proposes integrating cloud computing and CNN for
fire detection. The paper[17] lays out a new method of classifying Forest fire based on
the CNN model, called Fire_Net inspired by the AlexNet network, but improved with
15 layers, more efficient for the classification task. In [18], the authors propose a fire
detection deep learning model called ForestResNet, based on ResNet-50. Another
study in [19] applied a multi-layered classification model to forest fire detection.
Their method uses transition learning based on VGG-16, ResNet-50, and DenseNet-
121 to classify flames, smoke, fireless, and other objects in the image. The paper [20]
proposes the CNN model for smoke detection in Forest fires. Their proposed CNN
model applied batch normalization and multi-convolution to optimize and improve
the accuracy of classification.

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