Biometric task network
A selected biometric analysis task is performed in a deep neural network that includes a common feature extraction neural network, a plurality of task-specific neural networks, a segmentation neural network, a landmark mesh neural network, a plurality of soft target segmentation neural networks, and a plurality of expert pooling neural networks that perform the plurality of biometric analysis tasks by inputting the image to the common feature extraction network to determine latent variables. The latent variables can be input to the plurality of biometric task-specific neural networks to determine a plurality of first biometric analysis task outputs. The latent variables can be input to a landmark mesh neural network to determine a landmark mesh. The landmark mesh and the first biometric task outputs to a plurality of expert pooling neural networks to determine a plurality of second biometric task outputs.
1. A system, comprising:
a computer that includes a processor and a memory, the memory including instructions executable by the processor to provide output from a selected biometric analysis task that is one of a plurality of biometric analysis tasks, based on an image provided from an image sensor;
wherein the selected biometric analysis task is performed in a deep neural network that includes a common feature extraction neural network, a plurality of biometric task-specific neural networks, a segmentation neural network, a landmark mesh neural network, a plurality of soft target segmentation neural networks, and a plurality of expert pooling neural networks that perform the plurality of biometric analysis tasks by:
inputting the image to the common feature extraction network to determine latent variables;
inputting the latent variables to the plurality of biometric task-specific neural networks including emotion detection and head pose to determine a plurality of first biometric analysis task outputs;
inputting the latent variables to a landmark mesh neural network to determine a landmark mesh that includes polygons based on landmark locations that indicate features of a human face;
inputting the landmark mesh and the first biometric task outputs to a plurality of expert pooling neural networks to determine a plurality of second biometric task outputs; and
training the deep neural network by:
determining one or more first loss functions based on the plurality of second biometric task outputs;
combining the first loss functions to determine a joint loss function that gives more weight to loss functions for biometric tasks that include larger training datasets than biometric tasks that include smaller training datasets; and
backpropagating the one or more first loss function and the joint loss function back through the deep neural network.
2. The system of claim 1 , the instructions including further instructions to perform the plurality of biometric tasks by:
inputting the latent variables to the segmentation neural network to determine a first estimate segmentation map;
inputting the landmark mesh to the plurality of soft target segmentation neural networks to determine a plurality of soft target estimates; and
inputting the first estimate segmentation map and the plurality of soft target estimates to an expert pooling neural network to determine a segmentation prediction.
3. The system of claim 1 , wherein the output from the landmark mesh neural network is stored in a memory and processed to determine temporal data regarding one or more image segments.
4. The system of claim 1 , further comprising a device, wherein the instructions include instructions to operate the device based on the output from the deep neural network according to the selected biometric analysis task.
5. The system of claim 1 , wherein the biometric analysis tasks include biometric identification, liveliness, and facial segmentation.
6. The system of claim 1 , wherein the common feature extraction neural network includes a plurality of convolutional layers.
7. The system of claim 1 , wherein the plurality of task-specific neural networks includes a plurality of fully connected layers.
8. The system of claim 1 , wherein the segmentation neural network, the landmark mesh neural network, and the plurality of expert pooling neural networks include a plurality of fully connected layers.
9. The system of claim 1 , wherein the outputs from the plurality of task-specific neural networks are input to a Softmax function before being combined with the outputs from the common feature extraction neural network.
10. The system of claim 1 , wherein the outputs from the common feature extraction neural network are combined with the soft target estimates using expert pooling.
11. The system of claim 1 , wherein output from the expert pooling are input to a Softmax function before being combined to determine the first loss functions.
12. The system of claim 1 , wherein one or more outputs from the plurality of task-specific neural networks are set to zero during training.
13. A method, comprising:
providing output from a selected biometric analysis task that is one of a plurality of biometric analysis tasks, based on an image provided from an image sensor;
wherein the selected biometric analysis task is performed in a deep neural network that includes a common feature extraction neural network, a plurality of biometric task-specific neural networks, a segmentation neural network, a landmark mesh neural network, a plurality of soft target segmentation neural networks, and a plurality of expert pooling neural networks that perform the plurality of biometric analysis tasks by:
inputting the image to the common feature extraction network to determine latent variables;
inputting the latent variables to the plurality of biometric task-specific neural networks including emotion detection and head pose to determine a plurality of first biometric analysis task outputs;
inputting the latent variables to a landmark mesh neural network to determine a landmark mesh that includes polygons based on landmark locations that indicate features of a human face;
inputting the landmark mesh and the first biometric task outputs to a plurality of expert pooling neural networks to determine a plurality of second biometric task outputs;
training the deep neural network by:
determining one or more first loss functions based on the plurality of second biometric task outputs;
combining the first loss functions to determine a joint loss function that gives more weight to loss functions for biometric tasks that include larger training datasets than biometric tasks that include smaller training datasets; and
backpropagating the one or more first loss function and the joint loss function back through the deep neural network.
14. The method of claim 13 , further comprising:
inputting the latent variables to the segmentation neural network to determine a first estimate segmentation map;
inputting the landmark mesh to the plurality of soft target segmentation neural networks to determine a plurality of soft target estimates; and
inputting the first estimate segmentation map and the plurality of soft target estimates to an expert pooling neural network to determine a segmentation prediction.
15. The method of claim 13 , wherein the output from the landmark mesh neural network is stored in a memory and processed to determine temporal data regarding one or more image segments.
16. The method of claim 13 , further comprising a device, wherein the device is operated based on the output from the deep neural network according to the selected biometric analysis task.
17. The method of claim 13 , wherein the biometric analysis tasks include biometric identification, liveliness, and facial segmentation.
18. The method of claim 13 , wherein the common feature extraction neural network includes a plurality of convolutional layers.
19. The method of claim 13 , wherein the plurality of task-specific neural networks includes a plurality of fully connected layers.
20. The method of claim 13 , wherein the segmentation neural network, the landmark mesh neural network, and the plurality of expert pooling neural networks include a plurality of fully connected layers.