Scribd
Upload
6 views1 page

Senzori Incendiu-4

The document discusses different techniques for fire detection using deep learning and computer vision, including convolutional neural networks (CNN), You Only Look Once (YOLO) algorithms, and temporal modeling. It also describes the fire dataset collected and used for training models, containing over 4000 images of fires captured under various conditions.

Uploaded by

Cip Grecu
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
6 views1 page

Senzori Incendiu-4

The document discusses different techniques for fire detection using deep learning and computer vision, including convolutional neural networks (CNN), You Only Look Once (YOLO) algorithms, and temporal modeling. It also describes the fire dataset collected and used for training models, containing over 4000 images of fires captured under various conditions.

Uploaded by

Cip Grecu
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 1

Sensors 2023, 23, 7078 4 of 15

popular deep neural networks in deep learning, especially when it comes to computer
vision applications. It uses a special technique called convolution. In [12], detection of
fire and smoke through images and videos using deep learning algorithms is performed,
including a CNN-based architecture to train a model with many images for a dataset.
Dilated convolutional layers have been built to avoid depth of learning, which means to
learn larger data by ignoring the minute details. Classifying smoke and fire to reduce false
fire alarms is accomplished successfully in this research work. Differentiation between
smoke and fire in images and videos is accomplished by using a dilated CNN to learn the
robustness of features from the scene. The experimental results indicate that the proposed
method performs slightly better than well-known neural network architectures on their
custom dataset. However, the main limitation is that errors occur when pixel values come
closer to those of background and edges or are not detected by the CNN. Because it is a
custom-built dataset, it is computationally expensive. In [13] a deep learning-based fire
detection system called Detection and Temporal Accumulations (DTA) is used that imitates
the human detection process to improve the accuracy of fire detection while reducing
false detections and misinterpretations. The proposed method successfully interprets the
temporal SRoF behaviors and improves the fire detection accuracy. The faster R-CNN
model is used, which can detect multiple objects in a frame. It can detect fire, flames,
and smoke in a frame. Long short-term memory (LSTM) to accumulate the temporal
behaviors and to decide whether there is a fire or not in a short-term period is also used.
In [14], the authors designed a lightweight convolutional neural network for early fire and
smoke detection which successfully achieved competitive accuracy. It uses two satellite
imagery datasets and three smoke-related scene classes, namely, “Smoke”, “Clear”, and
“Other aerosol”. This model needs more improvements, as some of the smoke patches were
misinterpreted as “Clear” or “Other aerosol”, which is troubling for the early prediction of
fires [14].

2.3. Fire Detection Using YOLO (You Only Look Once) Algorithms
In [15], the YOLOV3 algorithm is used for a small-scale flame detection method.
This method was proposed to achieve the detection of different scales of flames using
an improved K-means clustering algorithm. The authors of [16] introduce a fire-YOLO
algorithm. It adds depth-separable convolution to YOLOv4 and helps to reduce the com-
putational costs of the model and improve the perceptual field of the feature layer by
using a cavity convolution method. The authors of [17] proposed a fire detection technique
for urban areas using ELASTIC-YOLOv3 as an improvement on YOLOv2 to amplify the
performance without introducing more parameters. Traditional fire algorithms, especially
the ones for nocturnal fire detection, suffered from issues like high light intensity, lack of
color information, changes in shapes and sizes of flames, etc., for which more advanced
and improved real-time fire detection and recognition systems came with modified YOLO
algorithms (v4, v5), as proposed in [18,19].

3. Proposed Work
3.1. Fire Dataset Description
To train the model, we collected a diverse range of images of fires from various internet
sources, including some videos to increase the size of dataset. For this purpose, images
obtained from distinct angles, focal lengths, and brightening conditions were utilized in our
dataset to elevate the accuracy of the system. Even after exploring different resources, the
images were not enough for the dataset, so additionally, images from publicly accessible
dataset platforms such as Roboflow and Kaggle were included to broaden the dataset, as
shown in Table 1. The illustration features fire items, flames, and burning displays. The
dataset’s diversity strengthens the model’s ability to generalize unseen or unexpected data
and adapt to changing conditions. Containing both diurnal and nocturnal images, our
dataset reached a total of 4186 images of fires (Figures 1 and 2).

You might also like