Continuous training methods for systems identifying anomalies in an image of an object
A system identifying anomalies in an image of an object is first trained using first sets of images corresponding to first anomaly types for the object. A model of the object is formed in a latent space. A label for each anomalous image is used to calculate vectors containing means and standard deviations for each first anomaly types. The means and standard deviations are used to calculate a log-likelihood loss for each first anomaly type. The system is retrained using second sets of images corresponding to second anomaly types for the object. The vectors are supplemented using labels for each second anomaly types. A statistically sufficient sample of information in the means and standard deviations vectors is supplied to the latent space. A log-likelihood loss for each of the first and second anomaly types is calculated based on their respective mean and standard deviation.
1. A computer-implemented continuous training method for a system identifying anomalies in an image of an object, comprising:
training the system by:
supplying, to an image encoder, one or more first sets of images corresponding to one or more first anomaly types for the object, the image encoder forming a model of the object in a latent space,
supplying labels to an anomaly encoder, each label corresponding to a respective image among the one or more first sets of images corresponding to the one or more first anomaly types for the object, each label identifying a related anomaly type for the object,
calculating, at the anomaly encoder, a vector containing a mean for each of one or more first model modes defined for the one or more first anomaly types,
calculating, at the anomaly encoder, a vector containing a standard deviation for each of the one or more first model modes defined for the one or more first anomaly types, and
calculating a log-likelihood loss for each of the one or more first anomaly types based on their respective mean and standard deviation; and
retraining the system by:
supplying, to the image encoder, one or more second sets of images corresponding to one or more second anomaly types for the object, the image encoder updating the model of the object in the latent space,
supplying additional labels, to the anomaly encoder, each additional label corresponding to a respective image among the one or more second sets of images corresponding to the one or more second anomaly types for the object, each additional label identifying a related anomaly type for the object,
updating, at the anomaly encoder, the vector containing the mean for each of the one or more first model modes defined for the one or more first anomaly types by adding a mean for each of one or more second model modes defined for the one or more second anomaly types,
updating, at the anomaly encoder, the vector containing the standard deviation for each of the one or more first model modes defined for the one or more first anomaly types by adding a standard deviation for each of one or more second model modes defined for the one or more second anomaly types,
supplying, to the latent space, a statistically sufficient sample of information contained in the vectors containing the means and standard deviations, and
calculating a log-likelihood loss for each of the first and second anomaly types based on their respective mean and standard deviation.
2. The method of claim 1 , wherein the model of the object is a flow-based model.
3. The method of claim 1 , wherein the model of the object is a generative adversarial network model.
4. The method of claim 1 , wherein the model of the object is a variational autoencoder model.
5. The method of claim 1 , wherein:
training the system further comprises using classification information for each of the one or more first anomaly types when forming the model of the object in the latent space; and
retraining the system further comprises using classification information for each of the one or more first anomaly types and for each of the one or more second anomaly types when updating the model of the object in the latent space.
6. The method of claim 5 , further comprising:
supplying, to a classifier, a first label for each image among the one or more first sets of images;
calculating, by the classifier, a first classification loss for each of the first anomaly types;
using the first classification losses for training the system;
supplying, to the classifier a second label for each image among the one or more second sets of images;
calculating, by the classifier, a second classification loss for each of the second anomaly types; and
using the second classification losses for retraining the system.
7. The method of claim 6 , further comprising:
supplying, to the classifier, a content of the latent space;
using, at the classifier, the content of the latent space to classify each of the one or more first anomaly types for the object; and
using, at the classifier, the content of the latent space to classify each of the one or more second anomaly types for the object.
8. The method of claim 7 , wherein the content of the latent space supplied to the classifier is a portion of the latent space.
9. The method of claim 1 , wherein the retraining of the system is performed without downtime of the system.
10. The method of claim 1 , wherein training the system further comprises:
supplying, to an image encoder the system, a set of anomaly-free images of an object;
encoding, by the image encoder, each anomaly free image of the object to form a corresponding image model in the latent space;
generating, in an image decoder of the system, an output image corresponding to each of the image models; and
calculating, in the system, a reconstruction loss based on a norm of differences between each anomaly-free image of the object and the corresponding output image.
11. The method of claim 10 , wherein the anomaly-free images of the object are augmented images.
12. The method of claim 1 , wherein training the system further comprises calculating a regularization loss based on a ratio of an output of a previous layer of the model of the object over an output of a current layer of the model of the object.
13. The method of claim 1 , wherein the images of the one or more first sets of images corresponding to the one or more first anomaly types for the object and the images of the one or more second sets of images corresponding to the one or more second anomaly types for the object are augmented images.
14. The method of claim 13 , wherein each augmented image is obtained by adding thereto an alteration selected from a random noise, a random cropping, a random rotation, a random set of white patches, a random set of black patches and a combination thereof.
15. A system for identifying anomalies in an object, comprising:
an image encoder;
an anomaly encoder; and
a training engine adapted to train the system by:
supplying, to the image encoder, one or more first sets of images corresponding to one or more first anomaly types for the object, the image encoder forming a model of the object in a latent space,
supplying labels to the anomaly encoder, each label corresponding to a respective image among the one or more first sets of images corresponding to the one or more first anomaly types for the object, each label identifying a related anomaly type for the object,
calculating, at the anomaly encoder, a vector containing a mean for each of one or more first model modes defined for the one or more first anomaly types,
calculating, at the anomaly encoder, a vector containing a standard deviation for each of the one or more first model modes defined for the one or more first anomaly types, and
calculating a log-likelihood loss for each of the one or more first anomaly types based on their respective mean and standard deviation; and
the training engine being also adapted to retrain the system by:
supplying, to the image encoder, one or more second sets of images corresponding to one or more second anomaly types for the object, the image encoder updating the model of the object in the latent space,
supplying additional labels to the anomaly encoder, each additional label corresponding to a respective image among the one or more second sets of images corresponding to the one or more second anomaly types for the object, each label identifying a related anomaly type for the object,
updating, at the anomaly encoder, the vector containing the mean for each of the one or more first model modes defined for the one or more first anomaly types by adding a mean for each of one or more second model modes defined for the one or more second anomaly types,
updating, at the anomaly encoder, the vector containing the standard deviation for each of the one or more first model modes defined for the one or more first anomaly types by adding a standard deviation for each of one or more second model modes defined for the one or more second anomaly types,
supplying, to the latent space, a statistically sufficient sample of information contained in the vectors containing the means and standard deviations, and
calculating a log-likelihood loss for each of the first and second anomaly types based on their respective mean and standard deviation.
16. The system of claim 15 , further comprising an image decoder, wherein:
the image encoder implements a first function;
the image decoder implements a second function, the second function being an inverse of the first function; and
the image encoder and the image decoder share a common set of weights.
17. The system of claim 16 , further comprising:
an input interface operatively connected to the image encoder; and
an output interface operatively connected to the image decoder;
wherein:
the input interface is adapted to receive an input image of the object from an image source and to provide the input image to the image encoder;
the image encoder is adapted to generate an image model based on the input image of the object;
the image decoder is adapted to generate a decoded image of the object based on the image model; and
the output interface is adapted to transmit the decoded image of the object to an image receiver.