Method and apparatus for rate-adaptive neural image compression with adversarial generators

1. A method of rate-adaptive neural image compression with adversarial generators, the method being performed by at least one processor, and the method comprising:

obtaining a first feature of an input image, using a first portion of a first neural network;

generating a first substitutional feature, based on the obtained first feature, using a second neural network;

encoding the generated first substitutional feature, using a second portion of the first neural network, to generate a first encoded representation;

compressing the generated first encoded representation;

decompressing the compressed representation; and

decoding the decompressed representation, using a third neural network, to reconstruct a first output image,

wherein the second neural network is trained based on a rate loss of the compressed representation and a distortion loss determined between the input image and the reconstructed first output image.

2. The method of claim 1 , wherein the second neural network is trained by:

encoding the obtained first feature, using a third portion of the first neural network, to generate a second encoded representation;

determining a representation discrimination loss between the generated first encoded representation and the generated second encoded representation, using a fourth neural network;

determining a feature discrimination loss between the generated first substitutional feature and the obtained first feature, using a fifth neural network; and

updating weight coefficients of the second neural network, the fourth neural network and the fifth neural network, to optimize the determined rate loss, the determined distortion loss, the determined representation discrimination loss and the determined feature discrimination loss.

3. The method of claim 1 , further comprising:

encoding the input image, using the first neural network, to generate the first encoded representation;

obtaining a second feature from the decompressed representation, using a first portion of the third neural network;

generating a second substitutional feature, based on the obtained second feature, using a fourth neural network; and

decoding the generated second substitutional feature, using a second portion of the third neural network, to reconstruct the first output image.

4. The method of claim 3 , wherein the fourth neural network is trained by:

decoding the obtained second feature, using a third portion of the third neural network, to reconstruct a second output image;

determining a representation discrimination loss between the reconstructed first output image and the reconstructed second output image, using a fifth neural network;

determining a feature discrimination loss between the generated second substitutional feature and the obtained second feature, using a sixth neural network; and

updating weight coefficients of the fourth neural network, the fifth neural network and the sixth neural network, to optimize the determined rate loss, the determined distortion loss, the determined representation discrimination loss and the determined feature discrimination loss.

5. The method of claim 3 , further comprising:

obtaining a third feature of the input image, using the first neural network; and

generating an attention map, based on the obtained third feature,

wherein the generating the second substitutional feature comprises generating the second substitutional feature, based on the obtained second feature and the generated attention map, using the fourth neural network.

6. The method of claim 3 , further comprising:

obtaining a third feature from the decompressed representation, using the first portion of the third neural network; and

generating an attention map, based on the obtained third feature,

wherein the generating the second substitutional feature comprises generating the second substitutional feature, based on the obtained second feature and the generated attention map, using the fourth neural network.

7. The method of claim 1 , further comprising:

obtaining a second feature of the input image, using the first portion of the first neural network; and

generating an attention map, based on the obtained second feature,

wherein the generating the first substitutional feature comprises generating the first substitutional feature, based on the obtained first feature and the generated attention map, using the second neural network.

8. An apparatus for rate-adaptive neural image compression with adversarial generators, the apparatus comprising:

at least one memory configured to store program code; and

at least one processor configured to read the program code and operate as instructed by the program code, the program code comprising:

first obtaining code configured to cause the at least one processor to obtain a first feature of an input image, using a first portion of a first neural network;

first generating code configured to cause the at least one processor to generate a first substitutional feature, based on the obtained first feature, using a second neural network;

first encoding code configured to cause the at least one processor to encode the generated first substitutional feature, using a second portion of the first neural network, to generate a first encoded representation;

compressing code configured to cause the at least one processor to compress the generated first encoded representation;

decompressing code configured to cause the at least one processor to decompress the compressed representation; and

first decoding code configured to cause the at least one processor to decode the decompressed representation, using a third neural network, to reconstruct a first output image,

wherein the second neural network is trained based on a rate loss of the compressed representation and a distortion loss determined between the input image and the reconstructed first output image.

9. The apparatus of claim 8 , wherein the second neural network is trained by:

encoding the obtained first feature, using a third portion of the first neural network, to generate a second encoded representation;

determining a representation discrimination loss between the generated first encoded representation and the generated second encoded representation, using a fourth neural network;

determining a feature discrimination loss between the generated first substitutional feature and the obtained first feature, using a fifth neural network; and

updating weight coefficients of the second neural network, the fourth neural network and the fifth neural network, to optimize the determined rate loss, the determined distortion loss, the determined representation discrimination loss and the determined feature discrimination loss.

10. The apparatus of claim 8 , wherein the program code further comprises:

second encoding code configured to cause the at least one processor to encode the input image, using the first neural network, to generate the first encoded representation;

second obtaining code configured to cause the at least one processor to obtain a second feature from the decompressed representation, using a first portion of the third neural network;

second generating code configured to cause the at least one processor to generate a second substitutional feature, based on the obtained second feature, using a fourth neural network; and

second decoding code configured to cause the at least one processor to decode the generated second substitutional feature, using a second portion of the third neural network, to reconstruct the first output image.

11. The apparatus of claim 10 , wherein the fourth neural network is trained by:

decoding the obtained second feature, using a third portion of the third neural network, to reconstruct a second output image;

determining a representation discrimination loss between the reconstructed first output image and the reconstructed second output image, using a fifth neural network;

determining a feature discrimination loss between the generated second substitutional feature and the obtained second feature, using a sixth neural network; and

updating weight coefficients of the fourth neural network, the fifth neural network and the sixth neural network, to optimize the determined rate loss, the determined distortion loss, the determined representation discrimination loss and the determined feature discrimination loss.

12. The apparatus of claim 10 , wherein the program code further comprises:

third obtaining code configured to cause the at least one processor to obtain a third feature of the input image, using the first neural network; and

third generating code configured to cause the at least one processor to generate an attention map, based on the obtained third feature,

wherein the second generating code is further configured to cause the at least one processor to generate the second substitutional feature, based on the obtained second feature and the generated attention map, using the fourth neural network.

13. The apparatus of claim 10 , wherein the program code further comprises:

third obtaining code configured to cause the at least one processor to obtain a third feature from the decompressed representation, using the first portion of the third neural network; and

third generating code configured to cause the at least one processor to generate an attention map, based on the obtained third feature,

wherein the second generating code is further configured to cause the at least one processor to generate the second substitutional feature comprises generating the second substitutional feature, based on the obtained second feature and the generated attention map, using the fourth neural network.

14. The apparatus of claim 8 , wherein the program code further comprises:

second obtaining code configured to cause the at least one processor to obtain a second feature of the input image, using the first portion of the first neural network; and

second generating code configured to cause the at least one processor to generate an attention map, based on the obtained second feature,

wherein the first generating code is further configured to cause the at least one processor to generate the first substitutional feature, based on the obtained first feature and the generated attention map, using the second neural network.

15. A non-transitory computer-readable medium storing instructions that, when executed by at least one processor for rate-adaptive neural image compression with adversarial generators, cause the at least one processor to:

obtain a first feature of an input image, using a first portion of a first neural network;

generate a first substitutional feature, based on the obtained first feature, using a second neural network;

encode the generated first substitutional feature, using a second portion of the first neural network, to generate a first encoded representation;

compress the generated first encoded representation;

decompress the compressed representation; and

decode the decompressed representation, using a third neural network, to reconstruct a first output image,

wherein the second neural network is trained based on a rate loss of the compressed representation and a distortion loss determined between the input image and the reconstructed first output image.

16. The non-transitory computer-readable medium of claim 15 , wherein the second neural network is trained by:

encoding the obtained first feature, using a third portion of the first neural network, to generate a second encoded representation;

determining a representation discrimination loss between the generated first encoded representation and the generated second encoded representation, using a fourth neural network;

determining a feature discrimination loss between the generated first substitutional feature and the obtained first feature, using a fifth neural network; and

updating weight coefficients of the second neural network, the fourth neural network and the fifth neural network, to optimize the determined rate loss, the determined distortion loss, the determined representation discrimination loss and the determined feature discrimination loss.

17. The non-transitory computer-readable medium of claim 15 , wherein the instructions, when executed by the at least one processor, further cause the at least one processor to:

encode the input image, using the first neural network, to generate the first encoded representation;

obtain a second feature from the decompressed representation, using a first portion of the third neural network;

generate a second substitutional feature, based on the obtained second feature, using a fourth neural network; and

decode the generated second substitutional feature, using a second portion of the third neural network, to reconstruct the first output image.

18. The non-transitory computer-readable medium of claim 17 , wherein the fourth neural network is trained by:

decoding the obtained second feature, using a third portion of the third neural network, to reconstruct a second output image;

determining a representation discrimination loss between the reconstructed first output image and the reconstructed second output image, using a fifth neural network;

determining a feature discrimination loss between the generated second substitutional feature and the obtained second feature, using a sixth neural network; and

updating weight coefficients of the fourth neural network, the fifth neural network and the sixth neural network, to optimize the determined rate loss, the determined distortion loss, the determined representation discrimination loss and the determined feature discrimination loss.

19. The non-transitory computer-readable medium of claim 17 , wherein the instructions, when executed by the at least one processor, further cause the at least one processor to:

obtain a third feature of the input image, using the first neural network;

generate an attention map, based on the obtained third feature; and

generate the second substitutional feature, based on the obtained second feature and the generated attention map, using the fourth neural network.

20. The non-transitory computer-readable medium of claim 17 , wherein the instructions, when executed by the at least one processor, further cause the at least one processor to:

obtain a third feature from the decompressed representation, using the first portion of the third neural network;

generate an attention map, based on the obtained third feature; and

generate the second substitutional feature, based on the obtained second feature and the generated attention map, using the fourth neural network.