202
6 Deep Learning and RFID System Physical Anti-Collision
successful applications of CNN in the fields of image feature extraction and recognition, it provides a new notion to solve the problem of image denoising [11, 12]. A
completely new convolutional neural network structure applied to image processing
is put forward, which can fully excavate the internal features and the prior knowledge
of the image. Plain neural networks can be easily trained by random gradient descent
on very large data sets. Secondly, the image processing results could be combined
with the subsequent deep belief network (DBN) to obtain the relationship between
the RFID tag group’s 3D positions and corresponding reading distance.
In order to optimize the tags’ geometrical distribution and improve the tags’
reading performance, this chapter introduces deep learning to pre-process RFID
multi-tag images and analyzes the reading performance of multi-tag structures in
depth. Therefore, the main contents of this chapter are as follows. First, the theory
related to deep learning is briefly reduced. Secondly, the constructed RFID multi-tag
3D measurement system based deep learning is introduced. Thirdly, the multi-level
wavelet CNN (MWCNN) and flexible feed-forward denoising convolutional neural
network (FDnCNN) are used to process images. The network denoises and analyzes
the multi-label network separately. Finally, the three-dimensional coordinates of the
multi-label and the corresponding reading distance are modeled by a deep belief
network (DBN), and the system is analyzed and evaluated.
6.1 RFID Multi-tag 3D Measurement System
6.1.1 System Architecture
The communication between readers and RFID multi-tag shares the same wireless
channel. When RFID multi-label enters the reader’s reading range, the data will
return to the reader simultaneously, arousing information collision. Data collision
will make the readers unable to identify the useful information of the labels, leading
to inaccurate label positioning, lower reading recognition efficiency, larger missed
reading rate, and longer recognition delay, which seriously limits the application field
of RFID. Therefore, the image processing of multi-label localization is proposed,
which not only can effectively avoid collision caused by channel interference, but
also can read and identify label information more quickly. It will undoubtedly be
a breakthrough direction. In [13, 14], an RFID multi-tag 2D measurement system
by one camera was designed to extract the position of the tags by flood filling and
morphological methods, and establish the direct linear transformation (DLT) between
the pixel coordinates and the actual spatial coordinates of the tags.
A 3D structure prediction system for the RFID tag group is proposed, which is to
find out the optimal geometric distribution structure of RFID tag groups and further
improve the reading distance of the RFID tag group. The schematic diagram of the
3D structure prediction system for the RFID tag group is shown in Fig. 6.1a The
physical map is shown in Fig. 6.1b. The block diagram is shown in Fig. 6.2. The
6 Deep Learning and RFID System Physical Anti-Collision
successful applications of CNN in the fields of image feature extraction and recognition, it provides a new notion to solve the problem of image denoising [11, 12]. A
completely new convolutional neural network structure applied to image processing
is put forward, which can fully excavate the internal features and the prior knowledge
of the image. Plain neural networks can be easily trained by random gradient descent
on very large data sets. Secondly, the image processing results could be combined
with the subsequent deep belief network (DBN) to obtain the relationship between
the RFID tag group’s 3D positions and corresponding reading distance.
In order to optimize the tags’ geometrical distribution and improve the tags’
reading performance, this chapter introduces deep learning to pre-process RFID
multi-tag images and analyzes the reading performance of multi-tag structures in
depth. Therefore, the main contents of this chapter are as follows. First, the theory
related to deep learning is briefly reduced. Secondly, the constructed RFID multi-tag
3D measurement system based deep learning is introduced. Thirdly, the multi-level
wavelet CNN (MWCNN) and flexible feed-forward denoising convolutional neural
network (FDnCNN) are used to process images. The network denoises and analyzes
the multi-label network separately. Finally, the three-dimensional coordinates of the
multi-label and the corresponding reading distance are modeled by a deep belief
network (DBN), and the system is analyzed and evaluated.
6.1 RFID Multi-tag 3D Measurement System
6.1.1 System Architecture
The communication between readers and RFID multi-tag shares the same wireless
channel. When RFID multi-label enters the reader’s reading range, the data will
return to the reader simultaneously, arousing information collision. Data collision
will make the readers unable to identify the useful information of the labels, leading
to inaccurate label positioning, lower reading recognition efficiency, larger missed
reading rate, and longer recognition delay, which seriously limits the application field
of RFID. Therefore, the image processing of multi-label localization is proposed,
which not only can effectively avoid collision caused by channel interference, but
also can read and identify label information more quickly. It will undoubtedly be
a breakthrough direction. In [13, 14], an RFID multi-tag 2D measurement system
by one camera was designed to extract the position of the tags by flood filling and
morphological methods, and establish the direct linear transformation (DLT) between
the pixel coordinates and the actual spatial coordinates of the tags.
A 3D structure prediction system for the RFID tag group is proposed, which is to
find out the optimal geometric distribution structure of RFID tag groups and further
improve the reading distance of the RFID tag group. The schematic diagram of the
3D structure prediction system for the RFID tag group is shown in Fig. 6.1a The
physical map is shown in Fig. 6.1b. The block diagram is shown in Fig. 6.2. The