Systems, methods, and apparatus for image classification with domain invariant regularization
A system and a method are disclosed for receiving an input image, using a domain invariant machine learning model to compute an output based on the input image, wherein the domain invariant machine learning model is trained using domain invariant regularization, and displaying information based on the output. Training using domain invariant regularization may include use of a loss function.
1 . A method comprising:
receiving an input image;
using a domain invariant machine learning model to compute an output based on the input image, wherein the domain invariant machine learning model is trained using domain invariant regularization comprising a style diversification loss function, wherein the domain invariant regularization uses a domain invariant loss function that accounts for a content-specific feature comprising subject information or category information and a domain-specific feature comprising specific style information of the input image;
displaying information based on the output; and
wherein the domain invariant machine learning model is trained to be invariant to domain-specific features using training images having disentangled content-specific and domain-specific feature spaces, the training images generated based on a multi-domain image-to-image translation network.
2 . The method of claim 1 , wherein the domain invariant regularization comprises using a domain invariant loss.
3 . The method of claim 1 , wherein the domain invariant machine learning model is trained using a first image of the training images, the first image comprising the content-specific feature and a first domain-specific feature and a second image of the training images, the second image comprising the content-specific feature and a second domain-specific feature.
4 . The method of claim 3 , wherein the second image is generated from the first image.
5 . The method of claim 3 , wherein the domain invariant machine learning model is trained by using a classification loss for the first image, using a domain invariant loss for the second image, and combining the classification loss and the domain invariant loss.
6 . The method of claim 3 , wherein the second image is generated using an image generator.
7 . The method of claim 6 , wherein the image generator is trained using a classifier model.
8 . A system comprising:
one or more processors;
a memory storing instructions which, when executed by the one or more processors, causes a domain invariant machine learning model to compute an output based on an input image, wherein the domain invariant machine learning model is trained using domain invariant regularization comprising a style diversification loss function and a loss function implemented using a distance function, wherein the domain invariant regularization uses a domain invariant loss function that accounts for a content-specific feature comprising subject information or category information and a domain-specific feature comprising specific style information of the input image; and
wherein the domain invariant machine learning model is trained to be invariant to domain-specific features using training images having disentangled content-specific and domain-specific feature spaces, the training images generated based on a multi-domain image-to-image translation network.
9 . The system of claim 8 , wherein the domain invariant regularization comprises using a domain invariant loss.
10 . The system of claim 8 , wherein the domain invariant machine learning model is trained using a first image of the training images, the first image comprising the content-specific feature and a first domain-specific feature and a second image of the training images, the second image comprising the content-specific feature and a second domain-specific feature.
11 . The system of claim 10 , wherein the second image is generated from the first image.
12 . The system of claim 10 , wherein the domain invariant machine learning model is trained by using a classification loss for the first image, using a domain invariant loss for the second image, and combining the classification loss and the domain invariant loss.
13 . The system of claim 10 , wherein the second image is generated using an image generator.
14 . The system of claim 13 , wherein the image generator is trained using a classifier model.
15 . A method of training a machine learning model, the method comprising:
receiving a plurality of input images;
generating a plurality of perturbed input images based on the plurality of input images, wherein the plurality of perturbed input images are generated based on a combination of a content-specific feature comprising subject information or category information and a domain-specific feature comprising specific style information of the plurality of input images;
training the machine learning model to be domain invariant using the plurality of input images, the plurality of perturbed input images, a style diversification loss function, and a loss function implemented using an expectation function; and
where in the machine learning model is a domain invariant machine learning model trained to be invariant to domain-specific features using training images having disentangled content-specific and domain-specific feature spaces, the training images generated based on a multi-domain image-to-image translation network.
16 . The method of claim 15 , wherein training the machine learning model to be domain invariant comprises using a domain invariant loss.
17 . The method of claim 16 , wherein training the machine learning model to be domain invariant comprises:
computing a classification loss for one of the plurality of input images;
computing the domain invariant loss for one of the plurality of perturbed input images; and
combining the classification loss and the domain invariant loss.
18 . The method of claim 15 , wherein one of the plurality of input images comprises the content-specific feature and a first domain-specific feature, and generating the plurality of perturbed input images comprises combining the content-specific feature with a second domain-specific feature.
19 . The method of claim 15 , wherein the plurality of perturbed input images are generated using an image generator.
20 . The method of claim 19 , wherein the image generator is trained using a classifier model.