Fusing multi-spectral images for identity authentication

1. A method performed by one or more data processing apparatus, the method comprising:

obtaining a multi-spectral image that depicts an eye of a user, wherein the multi-spectral image comprises a plurality of registered two-dimensional channels, and each two-dimensional channel corresponds to a different spectrum of the multi-spectral image;

processing the multi-spectral image using an encoder neural network to generate a fused image, wherein the fused image has a single two-dimensional channel;

determining a set of features characterizing the eye of the user from the fused image; and

determining an identity of the user based at least in part on the set of features characterizing the eye of the user, comprising:

using the set of features characterizing the eye of the user to determine a respective match score between the user and each of one or more reference users, wherein the match score between the user and a reference user characterizes a likelihood that the identity of the user matches an identity of the reference user; and

determining the identity of the user matches the identity of a particular reference user based on the match scores.

2. The method of claim 1 , wherein the multi-spectral image comprises one or more of: a red color channel, a green color channel, a blue color channel, an infrared channel, and an ultraviolet channel.

3. The method of claim 1 , wherein the encoder neural network comprises a plurality of convolutional neural network layers.

4. The method of claim 1 , wherein determining the set of features characterizing the eye of the user from the fused image comprises:

identifying positions of a plurality of interest points in the fused image; and

generating a respective feature descriptor corresponding to each interest point using the fused image.

5. The method of claim 1 , wherein determining a match score between the user and a reference user comprises:

determining a first match score based on eyeprint features that characterize a portion of the eye outside an iris of the eye;

determining a second match score based on iris features that characterize the iris of the eye; and

determining the match score based on the first match score and the second match score.

6. The method of claim 1 , wherein the encoder neural network is trained by a plurality of operations comprising:

obtaining a training multi-spectral image;

processing the training multi-spectral image using the encoder neural network, in accordance with current values of encoder neural network parameters, to generate a training fused image;

generating a target image, comprising:

segmenting the training multi-spectral image into multiple regions;

determining, for each region, a target region based on an appearance of the region in one or more channels of the training multi-spectral image; and

spatially concatenating the target regions to generate the target image;

determining gradients of an unsupervised loss with respect to the encoder neural network parameters, wherein the unsupervised loss comprises a measure of similarity between: (i) the training fused image, and (ii) the target image; and

using the gradients to adjust the current values of the encoder neural network parameters.

7. The method of claim 6 , wherein the multiple regions comprise one or more of: a skin region, a sclera region, and an iris region.

8. The method of claim 1 , wherein the encoder neural network is trained by a plurality of operations comprising:

obtaining a training multi-spectral image that depicts an eye of a training user;

processing the training multi-spectral image using the encoder neural network, in accordance with current values of encoder neural network parameters, to generate a training fused image;

processing the training fused image using a classification neural network, in accordance with current values of classification neural network parameters, to generate a classification output that characterizes a predicted identity of the training user;

determining gradients of a supervised loss with respect to the encoder neural network parameters, wherein the supervised loss is based on the classification output; and

using the gradients to adjust the current values of the encoder neural network parameters.

9. A system comprising:

one or more computers; and

one or more storage devices communicatively coupled to the one or more computers, wherein the one or more storage devices store instructions that, when executed by the one or more computers, cause the one or more computers to perform operations comprising:

obtaining a multi-spectral image that depicts an eye of a user, wherein the multi-spectral image comprises a plurality of registered two-dimensional channels, and each two-dimensional channel corresponds to a different spectrum of the multi-spectral image;

processing the multi-spectral image using an encoder neural network to generate a fused image, wherein the fused image has a single two-dimensional channel;

determining a set of features characterizing the eye of the user from the fused image; and

determining an identity of the user based at least in part on the set of features characterizing the eye of the user, comprising:

using the set of features characterizing the eye of the user to determine a respective match score between the user and each of one or more reference users, wherein the match score between the user and a reference user characterizes a likelihood that the identity of the user matches an identity of the reference user; and

determining the identity of the user matches the identity of a particular reference user based on the match scores.

10. The system of claim 9 , wherein the multi-spectral image comprises one or more of: a red color channel, a green color channel, a blue color channel, an infrared channel, and an ultraviolet channel.

11. The system of claim 9 , wherein the encoder neural network comprises a plurality of convolutional neural network layers.

12. The system of claim 9 , wherein determining the set of features characterizing the eye of the user from the fused image comprises:

identifying positions of a plurality of interest points in the fused image; and

generating a respective feature descriptor corresponding to each interest point using the fused image.

13. The system of claim 9 , wherein determining a match score between the user and a reference user comprises:

determining a first match score based on eyeprint features that characterize a portion of the eye outside an iris of the eye;

determining a second match score based on iris features that characterize the iris of the eye; and

determining the match score based on the first match score and the second match score.

14. One or more non-transitory computer storage media storing instructions that when executed by one or more computers cause the one or more computers to perform operations comprising:

obtaining a multi-spectral image that depicts an eye of a user, wherein the multi-spectral image comprises a plurality of registered two-dimensional channels, and each two-dimensional channel corresponds to a different spectrum of the multi-spectral image;

processing the multi-spectral image using an encoder neural network to generate a fused image, wherein the fused image has a single two-dimensional channel;

determining a set of features characterizing the eye of the user from the fused image; and

determining an identity of the user based at least in part on the set of features characterizing the eye of the user, comprising:

using the set of features characterizing the eye of the user to determine a respective match score between the user and each of one or more reference users, wherein the match score between the user and a reference user characterizes a likelihood that the identity of the user matches an identity of the reference user; and

determining the identity of the user matches the identity of a particular reference user based on the match scores.

15. The non-transitory computer storage media of claim 14 , wherein the multi-spectral image comprises one or more of: a red color channel, a green color channel, a blue color channel, an infrared channel, and an ultraviolet channel.

16. The non-transitory computer storage media of claim 14 , wherein the encoder neural network comprises a plurality of convolutional neural network layers.

17. The non-transitory computer storage media of claim 14 , wherein determining the set of features characterizing the eye of the user from the fused image comprises:

identifying positions of a plurality of interest points in the fused image; and

generating a respective feature descriptor corresponding to each interest point using the fused image.