6.2 The Theory RFID Multi-tag Image Denoising by FDnCNN
211
expense of removing the details of the image. In order to guarantee that M plays a
role in balance control, the convolution filter is standardized by orthogonalization
during training.
6.2.3 Loss Function
The input of FDnCNN is the noise image of RFID multi-label y = x + v. In image
denoising, the four sub-images and a noise grade map M are selected as the input of
the nonlinear network FDnCNN to continuously learn the relationship between the
potential and the estimated clean image in the network.
Such internal relations are expressed by parameters in the network. In the convolutional neural network, the network parameters are adjusted through the error backpropagation algorithm in continuous learning. Assuming that the original image is
x, the corresponding noise image is y, and the potential clean image is x
through
the networks, the relationship between the original image x and the estimated clean
image can be expressed as Eq. (6.6):
= (y, x,
∧
x; )
(6.6)
This is the loss function. The goal is to minimize the value of the loss function by
continuously training the network parameters as Eq. (6.7).
= min
(()
(6.7)
In general, in a discriminative learning model, the mean square error between
a potential clean image and an estimated clean image can be calculated as a loss
function for learning training parameters as shown in Eq. (6.8).
ζ(() =
1
2N
N
i=1
F(y i , M i ; ) − x i 2
(6.8)
In the training, Adam is the optimal algorithm to optimize the minimization loss
function. All hyper-parameters, beta1 and beta2, use the default values.
6.2.4 Experimental Result and Real-Time Analysis
(1) Implementation Details for Multi-label Image Denoising
In the image denoising of RFID multi-label dynamic localization system, the controlcomputer processor selected is Inter(R)Core(TM)i5-8300H CPU@2.3 GHz, the
211
expense of removing the details of the image. In order to guarantee that M plays a
role in balance control, the convolution filter is standardized by orthogonalization
during training.
6.2.3 Loss Function
The input of FDnCNN is the noise image of RFID multi-label y = x + v. In image
denoising, the four sub-images and a noise grade map M are selected as the input of
the nonlinear network FDnCNN to continuously learn the relationship between the
potential and the estimated clean image in the network.
Such internal relations are expressed by parameters in the network. In the convolutional neural network, the network parameters are adjusted through the error backpropagation algorithm in continuous learning. Assuming that the original image is
x, the corresponding noise image is y, and the potential clean image is x
through
the networks, the relationship between the original image x and the estimated clean
image can be expressed as Eq. (6.6):
= (y, x,
∧
x; )
(6.6)
This is the loss function. The goal is to minimize the value of the loss function by
continuously training the network parameters as Eq. (6.7).
= min
(()
(6.7)
In general, in a discriminative learning model, the mean square error between
a potential clean image and an estimated clean image can be calculated as a loss
function for learning training parameters as shown in Eq. (6.8).
ζ(() =
1
2N
N
i=1
F(y i , M i ; ) − x i 2
(6.8)
In the training, Adam is the optimal algorithm to optimize the minimization loss
function. All hyper-parameters, beta1 and beta2, use the default values.
6.2.4 Experimental Result and Real-Time Analysis
(1) Implementation Details for Multi-label Image Denoising
In the image denoising of RFID multi-label dynamic localization system, the controlcomputer processor selected is Inter(R)Core(TM)i5-8300H CPU@2.3 GHz, the