Page 87 - Kỷ yếu hội thảo khoa học lần thứ 12 - Công nghệ thông tin và Ứng dụng trong các lĩnh vực (CITA 2023)

71

Description of the Dataset. Of those 6420 images, there are 6300 images containing
Pistols and 120 images containing no Pistols, the purpose is to detect in cases where
there is no object to warn. In the 6300 there are 1240 images with difficult cases: Blurry
image, noise, low resolution, poor lighting, small objects, long distance, 1-point image.
- The dataset contains images of Pistols with various types, colors, designs, and ro-
tation angles of Pistol objects.
- 80% of data collected is artificial data (collected from public data sources) and 20%
of data is collected from reality.
- Each image can appear one or more objects (Pistols) with different sizes, all Pistols
must be labeled so each image can have more than 1 label.
- The dataset is divided by 70% for training data set and 30% for test and evaluation.

Fig. 7. Image of Pistol in training

Model Training Selection. In this step, we select a training model including:

Table 2. Training model

Model Version
YOLO V5 YOLOV5-n YOLOV5-m YOLOV5-l
YOLOV7 YOLOV7- x YOLOV7- w6 YOLOV7-E6
YOLOV8 YOLOV8- l YOLOV8- x

In this paper, we used the versions in Table 2 to perform the training, which are the
fastest and best versions for the YOLO series so far.

Installs and Experiments. The model training process is carried out using a
Pre-trained model as a basis, testing versions on the Google Colab platform with
specific configurations as follows:
CPU model: Intel ® Xeon ® CPU 2.20 GHz
Number of cores: 2
CPU frequency: 2200.158 MHz
Total size of Disk: 466.8 GB
Total amount of Memory: 12985 MB
System uptime: 0 days, 0 hour 1 min
Load average: 1.46, 0.51, 0.18
OS: Ubuntu 20.04.5 LTS
Arch: x86_64 (64 Bit)
Kernel: 5.10.147+

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