Neural network systems and methods for removing noise from signals

1. A method for removing noise from a data signal, the method comprising:

(a) receiving a first data signal that contains noise;

(b) removing, by a neural network, the noise from the first data signal to produce a processed data signal that represents the first data signal without the noise;

(c) calculating, from the first data signal and the processed data signal, a first noise signal representing the noise from the first data signal;

(d) generating a simulated noise signal;

(e) generating a comparison value by comparing the first noise signal and the simulated noise signal, the comparison value representing a correspondence between characteristics of the first noise signal and the simulated noise signal; and

(f) training the neural network using the comparison value to improve performance of the neural network in producing the processed data signal.

2. The method of claim 1 wherein generating the simulated noise signal comprises generating the simulated noise signal having characteristics corresponding to characteristics of the first noise signal.

3. The method of claim 1 wherein the first data signal represents an image having pixels, and generating the simulated noise signal comprises simulating noise by superimposing effects of multiple point scatterers in one or more pixels of the image.

4. The method of claim 1 wherein generating the simulated noise signal comprises simulating noise using a circularly complex Gaussian noise model.

5. The method of claim 1 wherein generating the simulated noise signal comprises applying a taper function to the simulated noise signal.

6. The method of claim 1 wherein training the neural network further includes training the neural network with one or more of: an adversarial loss value, a histogram loss value, a sum smooth loss value, or a smoothness loss value.

7. The method of claim 1 wherein the first data signal represents an image and calculating the first noise signal comprises calculating a difference between the first data signal and the processed data signal.

8. The method of claim 1 wherein:

the first data signal represents a radar image;

the noise contained in the first data signal comprises speckle; and

the simulated noise signal contains speckle.

9. The method of claim 1 wherein the step of generating the comparison value is performed by a second neural network.

10. The method of claim 1 wherein:

receiving the first data signal comprises receiving the first data signal from a system that produced the noise; and

generating the simulated noise signal comprises generating the simulated noise signal to have one or more noise characteristics in common with the system.

11. The method of claim 10 wherein the one or more noise characteristics include at least one of:

a frequency spectrum of the noise;

a power of the noise;

a signal to noise ratio; and

a randomness profile.

12. A system for removing noise from a data signal, the system comprising:

a first neural network configured to:

receive a first data signal that includes noise;

process the first data signal to produce a processed data signal representing the first data signal without the noise; and

a processor configured to:

generate a noise signal representing the noise included in the first data signal;

generate a simulated noise signal; and

compare the noise signal with the simulated noise signal to generate a comparison value that represents a correspondence between characteristics of the noise signal and characteristics of the simulated noise signal;

wherein the first neural network is trained with the comparison value to improve an ability of the first neural network to remove noise from data signals.

13. The system of claim 12 wherein the processor is configured to implement a second neural network, the second neural network configured to compare the noise signal with the simulated noise signal and generate the comparison value.

14. The system of claim 13 wherein the second neural network is a discriminatory neural network.

15. The system of claim 12 wherein the first neural network is further configured to produce a second data signal representing the first data signal without the noise.

16. The system of claim 12 wherein the processor is configured to generate the noise signal by calculating a difference between the first data signal and the processed data signal.

17. The system of claim 12 wherein the first data signal represents a radar image and the noise is speckle in the radar image.

18. The system of claim 12 wherein the first neural network is a self-supervised training network.

19. The system of claim 12 wherein the first neural network is a generative adversarial network.

20. The system of claim 12 wherein generating the simulated noise signal comprises using a circularly complex Gaussian noise model.

21. In a radar system, a method of training a neural network to remove speckle from radar images, the method comprising:

receiving data representing a radar image;

removing, by the neural network, speckle from the data representing the radar image to produce a despeckled image;

generating a speckle signal by subtracting the despeckled image from the radar image;

generating a simulated speckle signal;

comparing the speckle signal to the simulated speckle signal to generate a comparison value representing a correspondence between characteristics of the speckle signal and characteristics of the simulated speckle signal; and

training the neural network with the comparison value.

22. The method of claim 21 wherein the neural network is a generative adversarial network.

23. In a radar system, a method of despeckling a radar image, the method comprising:

receiving data representing a radar image;

processing, by a neural network, the received data to remove speckle from the radar image to produce processed data that represents the radar image without the speckle;

generating, from the received data and the processed data, a speckle signal representing the removed speckle;

generating a simulated speckle signal;

generating a comparison value by comparing the speckle signal to the simulated speckle signal; and

using the comparison value to train the neural network to improve an ability of the neural network to remove speckle from radar images.