Page 177 - Kaleidoscope Academic Conference Proceedings 2020

Industry-driven digital transformation

2.2.1 Primary screening 3.1 Generation of a suspicious object database via
simulation
In the preliminary screening, the visible light camera can
monitor each person in the surveillance area and record To establish a suspicious object database composed of various
their facial images and related information. At the same passive sensor images by simulation, two parameters should
time, the W-band radars can identify whether any person be considered. First, we should consider the simulation
carries a suspicious object (metal, etc.) 15 meters away. environment of the entire system, including temperature,
During this process, the suspicious object detection system reflection, blur, variation, and noise. Then, the parameters of
will automatically pair the millimeter-wave image and visible suspicious objects should be considered, including the type,
light image for the detected same person. If a suspicious size, rotation, transformation, etc. To this end, we generated
person is detected, the system will automatically track it and 44 kinds of bottles, 41 kinds of forks, 37 kinds of knives and
send the security staff the relevant information to conduct a 45 kinds of scissors through simulation, with a total of 10516,
secondary screening. 9799, 8843, and 10755 samples, respectively.

2.2.2 Secondary screening 3.2 Generation of a suspicious object database by means
of active/passive imagers
In the secondary screening, a hybrid imager composed of
active and passive imagers will be employed to identify the In addition to simulation, we also generate real experimental
detailed suspicious objects (knife, gun, scissors, etc.) the images for a suspicious object database through active or
suspicious person is holding. In this process, visible light passive millimeter-wave imaging. In this system, we built
cameras will also be used to help security staff identify a simple anechoic chamber using active/passive imagers
suspicious persons, and the generated visible light images to generate real image data. Active or passive imagers
will be associated with millimeter-wave images. Moreover, generate millimeter-wave experimental images by letting the
AI technology is used in this stage to increase the recognition experimenter carry different numbers and types of suspicious
probability for suspicious objects. objects (knives, simulated bombs, guns, liquids, phones, etc.).
During the experiment, the position of the experimenter and
2.3 AI-based suspicious object recognition technologies the direction of suspicious object changes all the time, which
is kept as consistent as possible with the actual scene. Usually,
In order to increase the probability of identifying suspicious suspicious objects are hidden in clothes or bags. In addition
objects in this system, we used AI technology to assist to this, we will also try to wrap suspicious objects with
this process, which is based on the developed suspicious different kinds of items (clothes, cotton, etc.). Throughout the
object database. Between different AI technologies, the experiment, 52 samples were generated by the active imager,
convolutional neural network (CNN) [4] is a representative and 1009 samples were generated by the passive imager.
deep-learning technology for image recognition and image
classification. Two critical features of CNN, which 3.3 Generation of a suspicious object database via GAN
make it special with other neural networks, are reducing
the computational complexity and ensuring translational 3.3.1 Fundamental of GAN
invariance. It mainly comprises two sections. The first
According to the description of [11], GAN was first proposed
section is used to extract the features and includes the
in [12]. It studies a two-player minimax game between
convolutional layer, pooling layer, batch normalization layer.
a generative network and a discriminative network .
The second section is in the same way as a neural network
Taking noisy sample I ∼ ?(I) (sampled from a normal or
works and used to make the classification, which includes
uniform distribution) as the input, the generative network
flatten layer, fully connected layer. In this paper, we directly
outputs new data (I), whose distribution ? 6 should be
use CNN technology so the performance of the AI part
close to that of the data distribution ? 30C0 . At the same
is determined by the CNN. Because the emphasis of this
time, the discriminative network is used to distinguish the
paper is to verify the feasibility of generating millimeter-wave
generated sample (I) ∼ ? 6 ( (I)). and the true data sample
images of suspicious objects for AI training through GAN and
G ∼ ? 30C0 (G). In the original GAN, this adversarial training
evaluate the factors that will affect the AI recognition rate,
process was expressed as
rather than the CNN itself, the technical details of CNN will
not be described here.
<8=<0G [;>6(1− ( (I)))] (1)
3. SUSPICIOUS OBJECT DATABASE TO SUPPORT G∼? 30C0 [;>6 (G)]+ G∼? I
AI-BASED RECOGNITION TECHNOLOGIES
3.3.2 Image generation via GAN
The AI-based suspicious object detection technology can be
employed to increase the recognition probability of suspicious GAN consists of two neural networks. In Figure 3, one of
objects in this system. To this end, there are three ways to the components is Generator, which as the name implies,
build a suspicious object database: simulation, active/passive generates data. The Generator inputs random noise (I)
imager, and GAN, which will be used for AI training. that corresponds to the characteristic seed of the generated

– 119 –