Radio wave signal receiver using convolutional neural network technology to improve signal to noise ratio
Systems, apparatuses and methods can provide for technology to perform a discrete cosine transform (DCT) on a first digital signal to obtain a frequency domain signal, wherein the frequency domain signal contains energy in a plurality of frequency bins, apply convolutional neural network (CNN) operations on the frequency domain signal, and output a probability for each of the plurality of frequency bins in accordance with the CNN operations, wherein each probability indicates whether the energy in a corresponding frequency bin contains a transmitted signal.
1 . A computing system comprising:
an analog to digital converter to convert a waveform to a first digital signal; and
a processor including logic coupled to one or more substrates, the logic to:
perform a discrete cosine transform (DCT) on the first digital signal to obtain a frequency domain signal, wherein the frequency domain signal contains energy in a plurality of frequency bins,
apply convolutional neural network (CNN) operations on the frequency domain signal,
output a probability for each of the plurality of frequency bins, wherein each probability indicates whether the energy in a corresponding frequency bin contains a transmitted signal,
identify one or more noisy frequency bins in the plurality of frequency bins, wherein the noisy frequency bins are associated with probabilities that are below a threshold,
suppress the energy in the noisy frequency bins, and
perform an inverse DCT on the frequency domain signal to obtain a second digital signal, wherein the second digital signal is to include energy that comprises the transmitted signal and the suppressed energy.
2 . The computing system of claim 1 , wherein the logic is further to:
obtain a plurality of one-dimensional (1D) segments from the frequency domain signal, and
combine a number of the 1D segments to obtain two-dimensional (2D) input segments, wherein the 2D input segments contain an additional channel dimension.
3 . The computing system of claim 2 , wherein the logic is further to iteratively perform, on a per network layer basis, a batch normalization operation on the 2D input segments, and wherein the batch normalization operation changes means and standard deviations of the 2D input segments.
4 . The computing system of claim 3 , wherein the logic is further to iteratively apply, on the per network layer basis, a nonlinear activation function to an output of the batch normalization operation.
5 . The computing system of claim 4 , wherein the logic is further to iteratively perform, on the per network layer basis, a 1D convolution operation on an output of the nonlinear activation function, and wherein an output of the 1D convolution operation is to describe intermediate features.
6 . The computing system of claim 1 , wherein the logic is further to bypass a classification of the energy in the plurality of frequency bins.
7 . A semiconductor apparatus comprising:
one or more substrates; and
logic coupled to the one or more substrates, wherein the logic is implemented at least partly in one or more of configurable or fixed-functionality hardware, the logic to:
perform a discrete cosine transform (DCT) on a first digital signal to obtain a frequency domain signal, wherein the frequency domain signal contains energy in a plurality of frequency bins;
apply convolutional neural network (CNN) operations on the frequency domain signal;
output a probability for each of the plurality of frequency bins in accordance with the CNN operations,
wherein each probability indicates whether the energy in a corresponding frequency bin contains a transmitted signal;
suppress the energy in one or more noisy frequency bins; and
perform an inverse DCT on the frequency domain signal to obtain a second digital signal,
wherein the second digital signal is to include energy that comprises the transmitted signal and the suppressed energy.
8 . The semiconductor apparatus of claim 7 , wherein the logic is further to:
identify the one or more noisy frequency bins in the plurality of frequency bins,
wherein the noisy frequency bins are associated with probabilities that are below a threshold.
9 . The semiconductor apparatus of claim 7 , wherein the logic is further to:
obtain a plurality of one-dimensional (1D) segments from the frequency domain signal; and
combine a number of the 1D segments to obtain two-dimensional (2D) input segments, wherein the 2D input segments contain an additional channel dimension.
10 . The semiconductor apparatus of claim 9 , wherein the logic is further to iteratively perform, on a per network layer basis, a batch normalization operation on the 2D input segments, and wherein the batch normalization operation is to change means and standard deviations of the 2D input segments.
11 . The semiconductor apparatus of claim 10 , wherein the logic is further to iteratively apply, on the per network layer basis, a nonlinear activation function to an output of the batch normalization operation.
12 . The semiconductor apparatus of claim 11 , wherein the logic is further to iteratively perform, on the per network layer basis, a 1D convolution operation on an output of the nonlinear activation function, and wherein an output of the 1D convolution operation is to describe intermediate features.
13 . The semiconductor apparatus of claim 7 , wherein the logic is further to bypass a classification of the energy in the plurality of frequency bins.
14 . At least one computer readable storage medium comprising a set of instructions, which when executed by a computing system, cause the computing system to:
perform a discrete cosine transform (DCT) on a first digital signal to obtain a frequency domain signal, wherein the frequency domain signal contains energy in a plurality of frequency bins;
apply convolutional neural network (CNN) operations on the frequency domain signal;
output a probability for each of the plurality of frequency bins in accordance with the CNN operations, wherein each probability indicates whether the energy in a corresponding frequency bin contains a transmitted signal;
suppress the energy in one or more noisy frequency bins; and
perform an inverse DCT on the frequency domain signal to obtain a second digital signal,
wherein the second digital signal is to include energy that comprises the transmitted signal and the suppressed energy.
15 . The at least one computer readable storage medium of claim 14 , wherein the instructions, when executed, further cause the computing system to:
identify the one or more noisy frequency bins in the plurality of frequency bins, wherein the noisy frequency bins are associated with probabilities that are below a threshold.
16 . The at least one computer readable storage medium of claim 14 ,
wherein the instructions, when executed, further cause the computing system to:
obtain a plurality of one-dimensional (1D) segments from the frequency domain signal; and
combine a number of the 1D segments to obtain two-dimensional (2D) input segments, wherein the 2D input segments contain an additional channel dimension.
17 . The at least one computer readable storage medium of claim 16 , wherein the instructions, when executed, further cause the computing system to iteratively perform, on a per network layer basis, a batch normalization operation on the 2D input segments, and wherein the batch normalization operation is to change means and standard deviations of the 2D input segments.
18 . The at least one computer readable storage medium of claim 17 , wherein the instructions, when executed, further cause the computing system to iteratively apply, on the per network layer basis, a nonlinear activation function to an output of the batch normalization operation.
19 . The at least one computer readable storage medium of claim 18 , wherein the instructions, when executed, further cause the computing system to iteratively perform, on the per network layer basis, a 1D convolution operation on an output of the nonlinear activation function, and wherein an output of the 1D convolution operation is to describe intermediate features.
20 . The at least one computer readable storage medium of claim 14 , wherein the instructions, when executed, further cause the computing system to bypass a classification of the energy in the plurality of frequency bins.