6.2 The Theory RFID Multi-tag Image Denoising by FDnCNN
215
(1) For the same σ , FDnCNN has the highest value of PSNR and SSIM. Hence,
FDnCNN has a better image denoising effect. When σ is small, the denoising
effect of FDnCNN, WNNM, and DnCNN is little different, but FDnCNN
should be at least 0.5 dB higher than others.
(2) As σ increases for the same method, the PSNR of FDnCNN decreases the
slowest, and the others decrease rapidly. Therefore, it can be seen that FDnCNN
has ascendant denoising stability obviously, and it achieves excellent results
in image denoising within a large noise grade range.
(3) At σ = 20, 30, it can be seen that the estimated clean images of FDnCNN can
achieve visual effects. WNNM denoising removes the image noise and image
details simultaneously. Compared with WNNM, DnCNN can get a favorable
denoising effect, and smooth the image. However, compared with the two ways,
FDnCNN can eliminate the noise in the flat areas, smooth the image structure,
and enhance the details of the multi-label edges. Consequently, the proposed
method is more conducive to multi-label localization and has the strongest
perception ability.
(4) Denoising of spatially variant AWGN
FDnCNN in handling spatially variant AWGN shows its flexibility. The spatially
variant AWGN is taken by AWGN with unit standard deviation is multiplied by the
noise level map with random Gauss variation, as shown in Fig. 6.11d. For the sake of
demonstrating the effectiveness of FDnCNN for spatially variant AWGN, compare
it with AWGN with a uniform noise level map, as shown in Fig. 6.11.
As we can see from Fig. 6.11, when dealing with AWGN with uniform noise, it
cannot remove the noise in the stronger area, but it can smooth the details of the area
with a lower noise level. When dealing with spatially variant AWGN, FDnCNN has
strong flexibility, and gets a good visual effect for different noise level areas.
(5) Real-time Performance
In addition to the analysis of visual quality and image quantification, a crucial indicator of evaluation is the speed of image denoising. All evaluation indicators are
performed in the MATLAB 2016a environment. In Table 6.2, real-time comparisons are made between WNNM, DnCNN, and FDnCNN, and the time shown is
the average time required for the corresponding procedures that run 50 times. The
time of WNNM is the average time to handle an image with a noise level of 10.
If the noise level is higher, WNNM takes a longer time. Besides, the cuDNN deep
learning library can speed up the calculation of DnCNN and FDnCNN. It takes about
two days to train the DnCNN network model, and FDnCNN was trained for about
three days. After the training is completed, the already trained model can always
handle the image denoising. For DnCNN, Table 6.2 represents the time required for
17 convolution layers to deal with a multi-label image. FDnCNN is the time required
for 15 convolution layers to denoise.
If we don’t take the CPU and GPU memory transfer time into account, both
DnCNN and FDnCNN will show breathtaking dominant positions over other
215
(1) For the same σ , FDnCNN has the highest value of PSNR and SSIM. Hence,
FDnCNN has a better image denoising effect. When σ is small, the denoising
effect of FDnCNN, WNNM, and DnCNN is little different, but FDnCNN
should be at least 0.5 dB higher than others.
(2) As σ increases for the same method, the PSNR of FDnCNN decreases the
slowest, and the others decrease rapidly. Therefore, it can be seen that FDnCNN
has ascendant denoising stability obviously, and it achieves excellent results
in image denoising within a large noise grade range.
(3) At σ = 20, 30, it can be seen that the estimated clean images of FDnCNN can
achieve visual effects. WNNM denoising removes the image noise and image
details simultaneously. Compared with WNNM, DnCNN can get a favorable
denoising effect, and smooth the image. However, compared with the two ways,
FDnCNN can eliminate the noise in the flat areas, smooth the image structure,
and enhance the details of the multi-label edges. Consequently, the proposed
method is more conducive to multi-label localization and has the strongest
perception ability.
(4) Denoising of spatially variant AWGN
FDnCNN in handling spatially variant AWGN shows its flexibility. The spatially
variant AWGN is taken by AWGN with unit standard deviation is multiplied by the
noise level map with random Gauss variation, as shown in Fig. 6.11d. For the sake of
demonstrating the effectiveness of FDnCNN for spatially variant AWGN, compare
it with AWGN with a uniform noise level map, as shown in Fig. 6.11.
As we can see from Fig. 6.11, when dealing with AWGN with uniform noise, it
cannot remove the noise in the stronger area, but it can smooth the details of the area
with a lower noise level. When dealing with spatially variant AWGN, FDnCNN has
strong flexibility, and gets a good visual effect for different noise level areas.
(5) Real-time Performance
In addition to the analysis of visual quality and image quantification, a crucial indicator of evaluation is the speed of image denoising. All evaluation indicators are
performed in the MATLAB 2016a environment. In Table 6.2, real-time comparisons are made between WNNM, DnCNN, and FDnCNN, and the time shown is
the average time required for the corresponding procedures that run 50 times. The
time of WNNM is the average time to handle an image with a noise level of 10.
If the noise level is higher, WNNM takes a longer time. Besides, the cuDNN deep
learning library can speed up the calculation of DnCNN and FDnCNN. It takes about
two days to train the DnCNN network model, and FDnCNN was trained for about
three days. After the training is completed, the already trained model can always
handle the image denoising. For DnCNN, Table 6.2 represents the time required for
17 convolution layers to deal with a multi-label image. FDnCNN is the time required
for 15 convolution layers to denoise.
If we don’t take the CPU and GPU memory transfer time into account, both
DnCNN and FDnCNN will show breathtaking dominant positions over other