212
6 Deep Learning and RFID System Physical Anti-Collision
operating system (OS) is Windows10 64-bit, and the GPU is NVIDIA Geforce
GTX 1050 Ti. All experimental results run in the software MATLAB 2016a, and
the proposed FDnCNN denoising model is trained with the MatconvNet toolbox
to implement convolutional neural networks (CNN) in computer vision applications. It is straightforward and efficient to train the most advanced deep convolution
neural networks. Meanwhile, MatconvNet runs on the support of the compiler Visual
Studio 2015. The version of the compute unified device architecture (CUDA) selects
NVIDIA CUDA Toolkit 9.0, which is a general parallel computing architecture that
enables the GPU to solve complex computational problems, and accelerate the operation. NVIDIA CuDNN9.1 is a GPU-accelerated library for deep neural networks,
which enables deep learning to calculate on the GPU. The MatconvNet package
supports software operation.
In the multi-label image denoiser based on FDnCNN, set the network depth to 15,
and the minimum loss function (Eq. (6.8)) is optimized by Adam. A small batch size
of 128 is selected. The learning rate of Adam algorithm starts from 10
–3 , until it is
reduced to 10
–4 . At the time, the training error stops decreasing. The small learning
rate is 10
–6 . The FDnCNN fine-tunes in an additional 50 iterations.
(2) Training and Testing Data Set
When multi-label images are denoised, the proposed method FDnCNN demands to be
trained firstly. We prepare an input–output pair of training data set {(y i , M i ; x i )}
N
i=1 ,
in which, M i is the noise level map. Forming the corresponding noise map M (in this
case, it is constant and all elements areσ) is added to the tidy patch x i by the AWGN
of σ[0, 100]. We used 400 Berkeley Segmentation Dataset (BSD) images and Set 68
as training data sets, randomly tailored N = 128 × 8000 patches for training, and the
patch size is set to 70 × 70 for grayscale images. The trained FDnCNN is embedded
in the control computer for receiving and handling RFID multi-label images acquired
by CCD. Thus, we used the multi-label image captured from the experiment as test
data sets.
(3) nComparisons to State-of-the-Arts for AWGN denoising
Due to its high flexibility, FDnCNN proposed in this paper to denoise image with
AWGN can ensure the compromise between image particular and denoising. In this
paper, PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index),
as two indexes of image quality evaluation, evaluates multi-label images at different
noise levels. At the different noise levels σ = 10, 30, 50, 70, the denoising results of
the proposed method are shown in Fig. 6.9.
It can be seen from Fig. 6.9:
(1) In a wide range of noise levels, FDnCNN can not only effectively denoise, but
also extract the edge portion of the labels from the visual perspective.
(2) Since the structure of the label images obtained by this system is relatively
simpler and has no complicated image particular, the value of PSNR is
comparatively higher.
6 Deep Learning and RFID System Physical Anti-Collision
operating system (OS) is Windows10 64-bit, and the GPU is NVIDIA Geforce
GTX 1050 Ti. All experimental results run in the software MATLAB 2016a, and
the proposed FDnCNN denoising model is trained with the MatconvNet toolbox
to implement convolutional neural networks (CNN) in computer vision applications. It is straightforward and efficient to train the most advanced deep convolution
neural networks. Meanwhile, MatconvNet runs on the support of the compiler Visual
Studio 2015. The version of the compute unified device architecture (CUDA) selects
NVIDIA CUDA Toolkit 9.0, which is a general parallel computing architecture that
enables the GPU to solve complex computational problems, and accelerate the operation. NVIDIA CuDNN9.1 is a GPU-accelerated library for deep neural networks,
which enables deep learning to calculate on the GPU. The MatconvNet package
supports software operation.
In the multi-label image denoiser based on FDnCNN, set the network depth to 15,
and the minimum loss function (Eq. (6.8)) is optimized by Adam. A small batch size
of 128 is selected. The learning rate of Adam algorithm starts from 10
–3 , until it is
reduced to 10
–4 . At the time, the training error stops decreasing. The small learning
rate is 10
–6 . The FDnCNN fine-tunes in an additional 50 iterations.
(2) Training and Testing Data Set
When multi-label images are denoised, the proposed method FDnCNN demands to be
trained firstly. We prepare an input–output pair of training data set {(y i , M i ; x i )}
N
i=1 ,
in which, M i is the noise level map. Forming the corresponding noise map M (in this
case, it is constant and all elements areσ) is added to the tidy patch x i by the AWGN
of σ[0, 100]. We used 400 Berkeley Segmentation Dataset (BSD) images and Set 68
as training data sets, randomly tailored N = 128 × 8000 patches for training, and the
patch size is set to 70 × 70 for grayscale images. The trained FDnCNN is embedded
in the control computer for receiving and handling RFID multi-label images acquired
by CCD. Thus, we used the multi-label image captured from the experiment as test
data sets.
(3) nComparisons to State-of-the-Arts for AWGN denoising
Due to its high flexibility, FDnCNN proposed in this paper to denoise image with
AWGN can ensure the compromise between image particular and denoising. In this
paper, PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index),
as two indexes of image quality evaluation, evaluates multi-label images at different
noise levels. At the different noise levels σ = 10, 30, 50, 70, the denoising results of
the proposed method are shown in Fig. 6.9.
It can be seen from Fig. 6.9:
(1) In a wide range of noise levels, FDnCNN can not only effectively denoise, but
also extract the edge portion of the labels from the visual perspective.
(2) Since the structure of the label images obtained by this system is relatively
simpler and has no complicated image particular, the value of PSNR is
comparatively higher.