Flat fine-grained image classification with progressive precision
Progressive precision image classifier and method of training include storing a dataset of labeled images, training a neural network to generate a classification vector comprising a plurality of confidence values, each confidence value corresponding to a classification, validating the trained neural network, calculating fine-grained confidence thresholds for each classification, wherein each classification represents a leaf-level classification in a hierarchical classification structure, and calculating coarse-level confidence thresholds for at least one parent class in the hierarchical classification structure, wherein each parent class defines a group of at least one leaf-level classification. Each label in the training data identifies a leaf-level classification in the hierarchical classification structure, and the classification vector includes a 1×N vector of confidence values, where N represents a number of leaf-level classifications output by the trained neural network. The neural network may be implemented as a convolution neural network with a single output head.
1. A system comprising:
a storage device configured to store a dataset of labeled images; and
a logic device configured to train a progressive precision image classifier by executing instructions comprising:
training a neural network to generate a classification vector comprising a plurality of confidence values, each confidence value corresponding to a classification;
validating the trained neural network;
calculating fine-grained confidence thresholds for each classification, wherein each classification represents a leaf-level classification in a hierarchical classification structure; and
calculating coarse-level confidence thresholds for at least one parent class in the hierarchical classification structure, wherein each parent class defines a group of at least one leaf-level classification.
2. The system of claim 1 , wherein the dataset of labeled images comprise a training dataset for use in training the neural network and a validation dataset for use in validating the trained neural network.
3. The system of claim 1 , wherein each label identifies a leaf-level classification in the hierarchical classification structure.
4. The system of claim 3 , wherein the classification vector comprising a 1×N vector of confidence values, where N represents a number of leaf-level classifications output by the trained neural network.
5. The system of claim 1 , wherein the neural network is a convolution neural network with a single output head.
6. The system of claim 1 , wherein validating the trained neural network comprises generating the classification vector for each validation input image; and storing top X confidence values, where X>=1, and a corresponding image label.
7. The system of claim 6 , wherein calculating the confidence thresholds for each leaf-level label comprises calculating a simple average, standard deviation, and/or harmonic mean of the confidence values corresponding to valid classifications associated with each label.
8. The system of claim 6 , wherein calculating coarse-level confidence thresholds for at least one parent class in the hierarchical classification structure, comprises analyzing the top X confidence values within the hierarchy for association with a parent class, and calculating the coarse-level confidence thresholds for at least one parent class using the confidence values of child classes of the parent class from the top X confidence values.
9. The system of claim 1 , further comprising distributing the trained progressive precision image classifier, including calculated confidence thresholds and the classification hierarchy for use in an inference system.
10. A method comprising:
providing a dataset of labeled images; and
training a progressive precision image classifier by executing instructions comprising:
training a neural network to generate a classification vector comprising a plurality of confidence values, each confidence value corresponding to a classification;
validating the trained neural network;
calculating fine-grained confidence thresholds for each classification, wherein each classification represents a leaf-level classification in a hierarchical classification structure; and
calculating coarse-level confidence thresholds for at least one parent class in the hierarchical classification structure, wherein each parent class defines a group of at least one leaf-level classification.
11. The method of claim 10 , wherein the dataset of labeled images comprise a training dataset for use in training the neural network and a validation dataset for use in validating the trained neural network.
12. The method of claim 10 , wherein each label identifies a leaf-level classification in the hierarchical classification structure.
13. The method of claim 12 , wherein the classification vector comprising a 1×N vector of confidence values, where N represents a number of leaf-level classifications output by the trained neural network.
14. The method of claim 10 , wherein the neural network is a convolution neural network with a single output head.
15. The method of claim 10 , wherein validating the trained neural network comprises generating the classification vector for each validation input image; and storing top X confidence values, where X>=1, and a corresponding image label.
16. The method of claim 15 , wherein calculating the confidence thresholds for each leaf-level label comprises calculating a simple average, standard deviation, and/or harmonic mean of the confidence values corresponding to valid classifications associated with each label.
17. The method of claim 15 , wherein calculating coarse-level confidence thresholds for at least one parent class in the hierarchical classification structure, comprises analyzing the top X confidence values within the hierarchy for association with a parent class, and calculating the coarse-level confidence thresholds for at least one parent class using the confidence values of child classes of the parent class from the top X confidence values.
18. The method of claim 10 , further comprising distributing the trained progressive precision image classifier, including calculated confidence thresholds and the classification hierarchy for use in an inference system.
19. The method of claim 10 , wherein the confidence threshold for each leaf-level label is calculated by averaging a group average confidence conference and class average confidence and subtracting a standard deviation for the group.
20. The method of claim 10 , wherein the classification vector comprises a plurality of sigmoid outputs.