Generative adversarial network-based target identification
A computing machine receives a real synthetic aperture radar (SAR) image including one or more targets. The real SAR image is one of a plurality of real SAR images in a training set. The computing machine generates, for the real SAR image, a model-based target shadow background (TSB) image using a three-dimensional (3D) model of the target. The computing machine generates, for the real SAR image and using an auto-encoder engine, an auto-encoder-generated TSB image using an artificial neural network (ANN). The computing machine computes, using a discriminator engine, an image difference between the auto-encoder-generated TSB image and the model-based TSB image. The computing machine adjusts weights in the auto-encoder engine based on the computed image difference.
1. A generative adversarial network (GAN) training apparatus comprising:
processing circuitry; and
a memory storing instructions which, when executed by the processing circuitry, cause the processing circuitry to perform operations comprising:
receiving a real synthetic aperture radar (SAR) image including a target, the real SAR image being one of a plurality of real SAR images in a training set;
generating, for the real SAR image, a model-based target shadow background (TSB) image using a three-dimensional (3D) model of the target;
generating, using the real SAR image as input to a neural network (NN) auto-encoder, an auto-encoder-generated TSB image;
computing, using a discriminator engine, an image difference between the auto-encoder-generated TSB image and the model-based TSB image; and
adjusting weights in the auto-encoder based on the computed image difference.
2. The apparatus of claim 1 , wherein each TSB image includes respective pixels identifying each pixel as being associated with the target, a shadow or a background.
3. The apparatus of claim 1 , the operations further comprising:
determining that the auto-encoder-generated TSB image and the model-based TSB image are not differentiable; and
determining that the auto-encoder is fully trained based on the auto-encoder-generated TSB image and the model-based TSB image not being differentiable for at least a threshold number of SAR images, including the real SAR image, from the training set.
4. The apparatus of claim 1 , the operations further comprising:
receiving, at the auto-encoder, a new SAR image;
generating, using the auto-encoder, a new TSB image for the new SAR image; and
identifying, using the new TSB image, one or more targets in the new SAR image.
5. The apparatus of claim 4 , wherein:
the auto-encoder comprises an encoder sub-engine, a latent vector, and a decoder sub-engine;
the encoder sub-engine receives the new SAR image and transmits data to the latent vector and to the decoder sub-engine via a skip connection;
the latent vector receives data from the encoder sub-engine and transmits processed data to the decoder sub-engine; and
the decoder sub-engine outputs the generated TSB image for the new SAR image.
6. The apparatus of claim 1 , wherein the 3D model comprises a voxel model.
7. The apparatus of claim 1 , wherein a GAN comprises the auto-encoder and the discriminator engine.
8. A non-transitory machine-readable medium storing instructions which, when executed by processing circuitry of one or more machines, cause the processing circuitry to perform operations comprising:
receiving a real synthetic aperture radar (SAR) image including a target, the real SAR image being one of a plurality of real SAR images in a training set;
generating, for the real SAR image, a model-based target shadow background (TSB) image using a three-dimensional (31)) model of the target;
generating, using the real SAR image as input to a neural network (NN) auto-encoder, an auto-encoder-generated TSB image;
computing, using a discriminator engine, an image difference between the auto-encoder-generated TSB image and the model-based TSB image; and
adjusting weights in the auto-encoder based on the computed image difference.
9. The machine-readable medium of claim 8 , wherein each TSB image includes respective pixels identifying each pixel as being associated with the target, a shadow or a background.
10. The machine-readable medium of claim 8 , the operations further comprising:
determining that the auto-encoder-generated TSB image and the model-based TSB image are not differentiable; and
determining that the auto-encoder is fully trained based on the auto-encoder-generated TSB image and the model-based TSB image not being differentiable for at least a threshold number of SAR images, including the real SAR image, from the training set.
11. The machine-readable medium of claim 8 , the operations further comprising:
receiving, at the auto-encoder, a new SAR image;
generating, using the auto-encoder, a new TSB image for the new SAR image; and
identifying, using the new TSB image, one or more targets in the new SAR image.
12. The machine-readable medium of claim 11 , wherein:
the auto-encoder engine comprises an encoder sub-engine, a latent vector, and a decoder sub-engine;
the encoder sub-engine receives the new SAR image and transmits data to the latent vector and to the decoder sub-engine via a skip connection;
the latent vector receives data from the encoder sub-engine and transmits processed data to the decoder sub-engine; and
the decoder sub-engine outputs the generated TSB image for the new SAR image.
13. The machine-readable medium of claim 8 , wherein the 3D model comprises a voxel model.
14. The machine-readable medium of claim 8 , wherein a generative adversarial network (GAN) comprises the auto-encoder and the discriminator engine.
15. A method, implemented at one or more computing machines, the method comprising:
receiving a real synthetic aperture radar (SAR) image including a target, the real SAR image being one of a plurality of real SAR images in a training set;
generating, for the real SAR image, a model-based target shadow background (TSB) image using a three-dimensional (3D) model of the target;
generating, using the real SAR image and a neural network (NN) auto-encoder, an auto-encoder-generated TSB image;
computing, using a discriminator engine, an image difference between the auto-encoder-generated TSB image and the model-based TSB image; and
adjusting weights in the auto-encoder based on the computed image difference.
16. The method of claim 15 , wherein each TSB image includes pixels identifying each pixel as being associated with the target, a shadow or a background.
17. The method of claim 15 , further comprising:
determining that the auto-encoder-generated TSB image and the model-based TSB image are not differentiable; and
determining that the auto-encoder is fully trained based on the auto-encoder-generated TSB image and the model-based TSB image not being differentiable for at least a threshold number of SAR images, including the real SAR image, from the training set.
18. The method of claim 15 , further comprising:
receiving, at the auto-encoder, a new SAR image;
generating, using the auto-encoder, a new TSB image for the new SAR image; and
identifying, using the new TSB image, one or more targets in the new SAR image.
19. The method of claim 18 , wherein:
the auto-encoder comprises an encoder sub-engine, a latent vector, and a decoder sub-engine;
the encoder sub-engine receives the new SAR image and transmits data to the latent vector and to the decoder sub-engine via a skip connection;
the latent vector receives data from the encoder sub-engine and transmits processed data to the decoder sub-engine; and
the decoder sub-engine outputs the generated TSB image for the new SAR image.
20. The method of claim 15 , wherein the 3D model comprises a voxel model.