End-to-end deep generative network for low bitrate image coding
A neural network system, a method and an apparatus for image compression are provided. The neural network may include a generator including an encoder, an entropy estimator, and a decoder, where the encoder receives an input image and generates an encoder output, a plurality of quantized feature entries are obtained based on the encoder output outputted at a last encoder block, the entropy estimator receives the plurality of quantized feature entries and calculates an entropy loss based on the plurality of quantized feature entries, and the decoder receives the plurality of quantized feature entries and generates a reconstructed image. Furthermore, the neural network may include a discriminator that determines whether the reconstructed image different from the input image based on a discriminator loss. Moreover, the generator may determine whether content of the reconstructed image matches content of the input image based on a generator loss including the entropy loss.
1. A neural network system implemented by one or more computers for compressing an image, comprising:
a generator comprising an encoder, an entropy estimator, and a decoder,
wherein the encoder receives an input image and generates an encoder output, a plurality of quantized feature entries are obtained based on the encoder output outputted at a last encoder block in the encoder, the entropy estimator receives the plurality of quantized feature entries and calculates an entropy loss based on the plurality of quantized feature entries, and the decoder receives the plurality of quantized feature entries and generates a reconstructed image; and
a discriminator that determines whether the reconstructed image is different from the input image based on a discriminator loss,
wherein a generator loss comprises the entropy loss and a combined content loss, and the combined content loss comprises a mean absolute error (MAE) and a multi-scale structural similarity index measure (MS-SSIM) between the reconstructed image and the input image, and
wherein the generator is configured to determine whether content of the reconstructed image matches content of the input image based on the generator loss.
2. The neural network system of claim 1 , wherein the plurality of quantized feature entries are obtained after performing soft quantization on the encoder output.
3. The neural network system of claim 1 , wherein the plurality of quantized feature entries are encoded to binary representations using a standardized binary arithmetic coding and sent to the decoder.
4. The neural network system of claim 1 , wherein entropy of the plurality of quantized feature entries is estimated based on a standard normal distribution.
5. The neural network system of claim 1 , wherein the generator loss further comprises an adversarial loss.
6. The neural network system of claim 1 , wherein the discriminator loss comprises a hinge loss that penalizes incorrect classification results.
7. The neural network system of claim 1 , wherein the reconstructed image is at a bitrate lower than a predetermined threshold.
8. The neural network system of claim 1 , wherein the discriminator updates discriminator network weights based on the discriminator loss, and the generator updates generator network weights based on the generator loss.
9. A method for compressing an image, comprising:
receiving, by an encoder in a generator in a neural network system, an input image and generating, by the encoder, an encoder output;
obtaining, by the encoder, a plurality of quantized feature entries based on the encoder output outputted at a last encoder block in the encoder;
receiving, by an entropy estimator in the generator, the plurality of quantized feature entries and calculating, by the entropy estimator, an entropy loss based on the plurality of quantized feature entries;
receiving, by a decoder in the generator, the plurality of quantized feature entries and generating, by the decoder, a reconstructed image;
determining, by a discriminator in the neural network system, whether the reconstructed image is different from the input image based on a discriminator loss, and
determining, by the generator, whether content of the reconstructed image matches content of the input image based on a generator loss, wherein the generator loss comprises the entropy loss and a combined content loss comprising a mean absolute error (MAE) and a multi-scale structural similarity index measure (MS-SSIM) between the reconstructed image and the input image.
10. The method of claim 9 , further comprising:
obtaining the plurality of quantized feature entries after performing soft quantization on the encoder output.
11. The method of claim 9 , further comprising:
encoding the plurality of quantized feature entries using a standardized binary arithmetic coding to binary representations and sending the plurality of quantized feature entries that are encoded to the decoder.
12. The method of claim 9 , further comprising:
estimating entropy of the plurality of quantized feature entries based on a standard normal distribution.
13. The method of claim 9 , wherein the generator loss further comprises an adversarial loss.
14. The method of claim 9 , further comprising:
calculating, by the discriminator, the discriminator loss comprising a hinge loss that penalizes incorrect classification results.
15. The method of claim 9 , wherein the reconstructed image is at a bitrate lower than a predetermined threshold.
16. The method of claim 9 , further comprising:
updating discriminator network weights based on the discriminator loss, or
updating generator network weights based on the generator loss.
17. An apparatus for compressing an image, comprising:
one or more processors; and
a memory configured to store instructions executable by the one or more processors, wherein the one or more processors, upon execution of the instructions, are configured to perform acts comprising:
receiving, by an encoder in a generator in a neural network system, an input image and generating, by the encoder, an encoder output;
obtaining, by the encoder, a plurality of quantized feature entries based on the encoder output outputted at a last encoder block in the encoder;
receiving, by an entropy estimator in the generator, the plurality of quantized feature entries and calculating, by the entropy estimator, an entropy loss based on the plurality of quantized feature entries;
receiving, by a decoder in the generator, the plurality of quantized feature entries and generating, by the decoder, a reconstructed image;
determining, by a discriminator in the neural network system, whether the reconstructed image is different from the input image based on a discriminator loss, and determining, by the generator, whether content of the reconstructed image matches the content of the original image based on a generator loss, wherein the generator loss comprises the entropy loss and a combined content loss comprising a mean absolute error (MAE) and a multi-scale structural similarity index measure (MS-SSIM) between the reconstructed image and the input image.
18. The apparatus of claim 17 , wherein the one or more processors are configured to perform acts further comprising:
wherein the discriminator loss comprises a hinge loss that penalizes incorrect classification results.