- |
-
Presenting a two-dimensional Restricted Boltzman Machine Network for Action Recognition in Video
Poster Presentation
Authors
Electrical and Computer Engineering University of Kashan, Kashan, Iran
Abstract
Action recognition is one of the most important
phases in understanding the video and the actions performed in it.
In recent years, different deep networks have been used for video
processing, including deep belief networks. This type of networks
has the same layers and fast learning ability. Deep belief networks
are composed of several layers of Restricted Boltzmann Machines,
and the similarity of these layers has led to the provision of
different types of deep belief networks for various applications.
The deep belief network will be the basic and constructive element
of the recurrent network implemented in this article. This article
has considered three important issues in the implementation of its
proposed method: First, since the video frames are two-dimensional, the network should be able to process them without
converting them into one-dimensional vectors. The second case is
to use limited Boltzmann machines to understand short-term time
concepts and the third case is to create the possibility of learning
long-term time concepts with recursive processing. We have
implemented the proposed method on two datasets, KTH and
UCF, and we have reached an accuracy of 94.17% and 91.86%,
respectively. The obtained results are also discussed.
phases in understanding the video and the actions performed in it.
In recent years, different deep networks have been used for video
processing, including deep belief networks. This type of networks
has the same layers and fast learning ability. Deep belief networks
are composed of several layers of Restricted Boltzmann Machines,
and the similarity of these layers has led to the provision of
different types of deep belief networks for various applications.
The deep belief network will be the basic and constructive element
of the recurrent network implemented in this article. This article
has considered three important issues in the implementation of its
proposed method: First, since the video frames are two-dimensional, the network should be able to process them without
converting them into one-dimensional vectors. The second case is
to use limited Boltzmann machines to understand short-term time
concepts and the third case is to create the possibility of learning
long-term time concepts with recursive processing. We have
implemented the proposed method on two datasets, KTH and
UCF, and we have reached an accuracy of 94.17% and 91.86%,
respectively. The obtained results are also discussed.
Keywords
Deep Learning; deep belief networks; restricted Boltzmann machine; action recognition; recurrent networks
Proceeding Title [Persian]
Presenting a two-dimensional Restricted Boltzman Machine Network for Action Recognition in Video
Authors [Persian]
Abstract [Persian]
Action recognition is one of the most important
phases in understanding the video and the actions performed in it.
In recent years, different deep networks have been used for video
processing, including deep belief networks. This type of networks
has the same layers and fast learning ability. Deep belief networks
are composed of several layers of Restricted Boltzmann Machines,
and the similarity of these layers has led to the provision of
different types of deep belief networks for various applications.
The deep belief network will be the basic and constructive element
of the recurrent network implemented in this article. This article
has considered three important issues in the implementation of its
proposed method: First, since the video frames are two-dimensional, the network should be able to process them without
converting them into one-dimensional vectors. The second case is
to use limited Boltzmann machines to understand short-term time
concepts and the third case is to create the possibility of learning
long-term time concepts with recursive processing. We have
implemented the proposed method on two datasets, KTH and
UCF, and we have reached an accuracy of 94.17% and 91.86%,
respectively. The obtained results are also discussed.
phases in understanding the video and the actions performed in it.
In recent years, different deep networks have been used for video
processing, including deep belief networks. This type of networks
has the same layers and fast learning ability. Deep belief networks
are composed of several layers of Restricted Boltzmann Machines,
and the similarity of these layers has led to the provision of
different types of deep belief networks for various applications.
The deep belief network will be the basic and constructive element
of the recurrent network implemented in this article. This article
has considered three important issues in the implementation of its
proposed method: First, since the video frames are two-dimensional, the network should be able to process them without
converting them into one-dimensional vectors. The second case is
to use limited Boltzmann machines to understand short-term time
concepts and the third case is to create the possibility of learning
long-term time concepts with recursive processing. We have
implemented the proposed method on two datasets, KTH and
UCF, and we have reached an accuracy of 94.17% and 91.86%,
respectively. The obtained results are also discussed.
Keywords [Persian]
Deep Learning، deep belief networks، restricted Boltzmann machine، action recognition، recurrent networks