Page 180 - Kaleidoscope Academic Conference Proceedings 2020

P. 180

2020 ITU Kaleidoscope Academic Conference

Table 3 – Object recognition results when epoch=5 (GAN training, gun: 43, knife: 43, other: 43, scissor: 43)

Evaluation results
Gun Knife Other Scissor Total Recall Average recall
Gun 42 1 14 1 58 0.724138
Knife 5 193 33 54 285 0.677193
Input 0.6867717
Other 5 4 60 0 69 0.869565
Scissor 1 5 5 10 21 0.47619
Total 53 203 112 65 433
Precision 0.792453 0.950739 0.535714 0.153846 Accuracy rate
Average Precision 0.608188 0.704388

Figure 9 – ROC curve
Figure 7 – Extend for CASE 1 5. STANDARDIZATION

Currently, AI technologies have been successfully applied
in ﬁelds such as ﬁnance, security, customer service,
and other industries. Standardization has a primary,
supporting, and leading role in the development of AI
and its industry. In recent years, diﬀerent standardization
organization has been studying AI issues and standardizing
related technologies.In the early stage, ISO/IEC JTC 1 has
carried out relevant standardization in key areas of AI
such as artiﬁcial intelligence vocabulary, human-computer
interaction, biometric features recognition, computer
image processing, and corresponding areas supported by
AI technologies such as cloud computing, big data,
and sensor networks. ISO has mainly carried out
Figure 8 – Extend for CASE 2 AI standardization research in industrial robots (ISO
11593:1996, ISO 9946:1999, ISO 14539:2000, ISO
for CNN training respectively. From the ﬁgures, we can
9787:1999, ISO 8373:2012), smart ﬁnance (ISO 19092:2008,
conclude that the accuracy rate of GAN images is higher
ISO 14742:2010, ISO 19038:2005), and smart driving
when the real experimental images used for CNN training
(ISO/TC 22 is responsible for formulating basic standards
as well as the CNN epoch number are small. It is more
related to road vehicles, and is conducting research on
accurate to use original images for CNN training with the
standardization of intelligent connected vehicles). IEC has
CNN epoch number increasing. Compared with these two
mainly carried out artiﬁcial intelligence standardization in the
ﬁgures, with more original images for CNN training (Figure
ﬁeld of wearable devices (IEC TC100 and IEC TC124). ITU
7), the above mentioned tendency becomes more remarkable
has worked on the development of AI standards since 2016.
and the accuracy of original images will faster exceed the
ITU-T has proposed draft proposals for AI, including ITU-T
GAN images (e.g. less CNN epoch number). Moreover, Y.AI4SC (Artiﬁcial Intelligence and IoT) and ITU-T Y.qos-ml
the ROC curve calculated based on Figure 8 was used to (Requirements of machine-learning-based QoS assurance),
evaluate the system. From this ﬁgure, we can get the same etc.
conclusion with Figure 8 when epoch is 50. Both GAN
images and original images, which are used for CNN training,
6. CONCLUSION
will achieve an excellent recognition rate. In general, the
GAN images are better than original images when epoch
In this work, an AI-based W-band suspicious object detection
number is 50.
system for moving persons are introduced. By employing
– 122 –

175 176 177 178 179 180 181 182 183 184 185