Vol.:(0123456789) 1 3
Journal of Ambient Intelligence and Humanized Computing
https://doi.org/10.1007/s12652-021-03261-2
ORIGINAL RESEARCH
An edge detection–based eGAN model for connectivity in ambient
intelligence environments
Cho Youn Lee
1
· Jin Gon Shon
1
· Ji Su Park
2
Received: 10 January 2021 / Accepted: 29 March 2021
© The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature 2021
Abstract
In generative adversarial networks (GANs), a generator network and discriminator network compete in deep learning tasks
to generate real images. To reduce the diference between the generated image and actual image, an edge GAN (eGAN)
model using edge detection was proposed. This eGAN model can utilize ambient intelligence, a human-centered technology
that includes IoT, smart cities, and autonomous driving. Ambient intelligence is essential for the interconnection between
humans and objects. The eGAN model was used to make this connectivity more accurate and reliable. Edge detection is an
edge feature that extracts the boundaries of an image and generate images in a fast manner; however, because its threshold is
arbitrarily set, the connectivity may be unstable. To solve this problem and improve the performance of the eGAN model, we
analyzed various GAN models and edge detection methods and proposed a new edge detection technology using threshold
settings. This edge detection method sets the threshold value for images, thereby increasing the accuracy of edge connection
and reducing the loss error between the image generated by the eGAN model and actual image. To evaluate the performance
of the eGAN model, the error between the generated image and actual image was compared by applying the GAN and eGAN
models to the same image dataset. Consequently, it was found that the performance of the eGAN model improved by 21%
in comparison to the existing GAN model.
Keywords Generative adversarial networks · Edge detection · EGAN · Threshold · Smart cities · Ambient intelligence
1 Introduction
In generative adversarial networks (GANs), a generator
network and discriminator network compete hostilely in
the feld of deep learning and generate image data similar
to real-world data. Initially, GAN becomes unstable as it
changes to various images in the learning process, which
creates realistic images. Theoretically, convergence is
guaranteed, but when minimization and maximization are
applied while learning the distribution of training data,
the theoretical assumption is broken, resulting in unstable
results. Because learning is unstable, images are repeatedly
generated using the generator and discriminator; this leads to
excessive crushing of the outline and poor image quality. To
improve this situation, various types of GAN models have
been actively examined and their utilization has increased as
a model that generates images in the feld of deep learning.
Because numerous images are composed of linear and non-
linear images, it is important to use edge detection methods
to create realistic similar images. GAN-based edge detection
is being actively studied to ensure accurate image data and
improve the super resolution (SR). GAN uses a loss function
to reduce the data error between the generated image and
actual image. In this study, an edge GAN (eGAN) model
using edge detection is proposed to reduce the loss error.
Edge detection is used to obtain useful information in com-
puter vision systems (Goodfellow 2014).
The edge detection process simplifes the analysis of
images while signifcantly reducing the amount of data to be
processed; moreover, it reveals useful structural information
about object boundaries. The frst and most obvious criterion
* Ji Su Park
jisupark@jj.ac.kr
Cho Youn Lee
qecche77@knou.ac.kr
Jin Gon Shon
jgshon@knou.ac.kr
1
Department of Computer Science, Graduate School, Korea
National Open University, Seoul, South Korea
2
Department of Computer Science and Engineering, Jeonju
University, Jeonju, South Korea