Page 86 - Kỷ yếu hội thảo khoa học lần thứ 12 - Công nghệ thông tin và Ứng dụng trong các lĩnh vực (CITA 2023)

70

a. In case of 1-point b. In case the object is partially covered
Fig. 6. Weapons are partially or completely covered

+ When using weapons to commit crimes, the subjects often perform acts at night,
wear black clothes, use black weapons, if the camera looks from a distance, it is very
difficult to identify the exact location of the weapon.
+ Pistols often have many common colors such as black, gold, silver, copper,... The
shape of the Pistol is almost like a hammer and is changed at many angles based on the
vertical, horizontal, and diagonal rotation. The shape also changes in pistol ratio due to
camera rotation, viewing angle, thereby causing a certain deviation.
- All of these problems are major challenges for weapon detection in different
circumstances.

Therefore, it is necessary to build a diverse learning dataset of objects in many
situations and circumstances, improve accuracy in difficult cases: distance; blurred,
noisy images; partially concealed; variety of species.

Data Collection and Pre-Processing. The dataset is divided into 3 data sets including
training set, test set and evaluation set, the images are then labeled to identify the
location of the pistol. Proceed to configure the parameters for the Model, after the

with the input of pistol images is taken from available images, videos, and actual
cam-
screen, and save the image or video.
The data set used for training and testing is detailed in the following table:

Table 1. Data (Pistol Data)

Label Training Testing Val
Pistol 3000 300 300

The experimental data set includes 6420 images with size 224x224, built by
extracting images from public sources with Pistol Detection Dataset [12], Pistol
classification dataset [13], Sohas weapon dataset [14]. The images in the dataset have
many different Pistol views and sizes.

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