Automatically updating a card-scan machine learning model based on predicting card characters

1. A system comprising:

at least one processor; and

a non-transitory computer readable medium comprising instructions that, when executed by the at least one processor, cause the system to:

capture, via a client device, a set of card images of a character-bearing card;

generate predicted card characters from the set of card images based at least on parameters of a card-scan machine learning model;

generate a card-scan gradient comprising multidimensional values representing one or more corrections to the predicted card characters;

determine an aggregated-classification-uncertainty metric reflecting a measure of confidence that the predicted card characters for the card-scan gradient are correct;

determine an aggregated-localization-tightness metric for the card-scan gradient representing a tightness of bounding boxes enclosing the predicted card characters;

determine an active-learning metric that indicates a predicted measure of accuracy improvement for the card-scan machine learning model based on the aggregated-classification-uncertainty metric and the aggregated-localization-tightness metric; and

update the card-scan machine learning model by modifying the card-scan machine learning model based on the active-learning metric.

2. The system of claim 1 , further comprising instructions that, when executed by the at least one processor, cause the system to determine the aggregated-classification-uncertainty metric by combining two or more classification-uncertainty metrics associated with the predicted card characters.

3. The system of claim 1 , further comprising instructions that, when executed by the at least one processor, cause the system to determine a classification-uncertainty metric as part of the aggregated-classification-uncertainty metric by comparing a first probability of a predicted card character corresponding to a first character and a second probability of the predicted card character corresponding to a second character.

4. The system of claim 1 , further comprising instructions that, when executed by the at least one processor, cause the system to generate the active-learning metric by:

determining a learning-modification weight to balance an effect of the aggregated-classification-uncertainty metric or the aggregated-localization-tightness metric on the active-learning metric; and

generating the active-learning metric based on a function comprising the aggregated-classification-uncertainty metric, the learning-modification weight, and the aggregated-localization-tightness metric.

5. The system of claim 1 , further comprising instructions that, when executed by the at least one processor, cause the system to:

rank the card-scan gradient relative to other card-scan gradients based on their respective active-learning metrics;

based on a ranking of card-scan gradients, identify the card-scan gradient as comprising low-information entropy or high-information entropy; and

label characters corresponding to the card-scan gradient to indicate correct card characters by:

providing a low-information-entropy card-scan gradient to an automatic labeling system for labeling; or

providing a high-information-entropy card-scan gradient for visualization within a labeling interface for labeling.

6. The system of claim 1 , further comprising instructions that, when executed by the at least one processor, cause the system to update the card-scan machine learning model by modifying the parameters of the card-scan machine learning model through retraining a stored version of the card-scan machine learning model to generate predictions of card characters based on the card-scan gradient and one or more correctly labeled card-scan gradients.

7. The system of claim 1 , further comprising instructions that, when executed by the at least one processor, cause the system to:

detect three-dimensional movement of the character-bearing card from the set of card images utilizing an optical-field detector of the card-scan machine learning model; and

based on the three-dimensional movement, verify that the character-bearing card associated with the set of card images is a physical character-bearing card.

8. A method comprising:

receiving, from a client device, a card-scan gradient comprising multidimensional values representing one or more corrections to predicted card characters made by a card-scan machine learning model for a character-bearing card;

determining an aggregated-classification-uncertainty metric reflecting a measure of confidence that the predicted card characters for the card-scan gradient are correct;

determining an aggregated-localization-tightness metric for the card-scan gradient representing a tightness of bounding boxes enclosing the predicted card characters;

determining an active-learning metric that indicates a predicted measure of accuracy improvement for the card-scan machine learning model based on the aggregated-classification-uncertainty metric and the aggregated-localization-tightness metric;

updating the card-scan machine learning model by modifying a stored version of the card-scan machine learning model based on the active-learning metric; and

providing the updated card-scan machine learning model to the client device.

9. The method of claim 8 , further comprising generating, from the card-scan gradient, a plurality of modified card-scan gradients comprising a first modified card-scan gradient and a second modified card-scan gradient by modifying one or more features of the card-scan gradient according to a probability distribution.

10. The method of claim 8 , wherein determining the aggregated-classification-uncertainty metric comprises combining two or more classification-uncertainty metrics associated with the predicted card characters.

11. The method of claim 9 , further comprising:

ranking the plurality of modified card-scan gradients relative to each other based on their respective active-learning metrics; and

based on the ranking, identifying the first modified card-scan gradient as comprising low-information entropy and the second modified card-scan gradient as comprising high-information-entropy.

12. The method of claim 11 , further comprising labeling characters corresponding to the first modified card-scan gradient and the second modified card-scan gradient to indicate correct card characters by:

providing the first modified card-scan gradient to an automatic labeling system for labeling; and

providing the second modified card-scan gradient for visualization within a labeling interface for labeling.

13. The method of claim 9 , wherein updating the card-scan machine learning model comprises retraining the stored version of the card-scan machine learning model to generate predictions of card characters based on the first modified card-scan gradient and the second modified card-scan gradient.

14. The method of claim 8 , further comprising providing, for display on the client device, a three-dimensional-card-scan interface comprising an animated guide depicting angle changes to a graphical card to instruct a user to rotate the character-bearing card in accordance with the animated guide.

15. A non-transitory computer readable medium comprising instructions that, when executed by at least one processor, cause a computing device to:

receive, from a client device, a card-scan gradient comprising multidimensional values representing one or more corrections to predicted card characters made by a card-scan machine learning model for a character-bearing card;

determine an aggregated-classification-uncertainty metric reflecting a measure of confidence that the predicted card characters for the card-scan gradient are correct;

determine an aggregated-localization-tightness metric for the card-scan gradient representing a tightness of bounding boxes enclosing the predicted card characters;

determine an active-learning metric that indicates a predicted measure of accuracy improvement for the card-scan machine learning model based on the aggregated-classification-uncertainty metric and the aggregated-localization-tightness metric;

update the card-scan machine learning model by modifying a stored version of the card-scan machine learning model based on the active-learning metric; and

provide the updated card-scan machine learning model to the client device.

16. The non-transitory computer readable medium of claim 15 , further comprising instructions that, when executed by the at least one processor, cause the computing device to determine the aggregated-classification-uncertainty metric by combining two or more classification-uncertainty metrics associated with the predicted card characters.

17. The non-transitory computer readable medium of claim 15 , wherein generating the active-learning metric comprises:

determining, based on a previous active-learning metric, a learning-modification weight to balance an effect of the aggregated-classification-uncertainty metric or the aggregated-localization-tightness metric on the active-learning metric; and

generate the active-learning metric based on a function comprising the aggregated-classification-uncertainty metric, the learning-modification weight, and the aggregated-localization-tightness metric.

18. The non-transitory computer readable medium of claim 15 , further comprising instructions that, when executed by the at least one processor, cause the computing device to:

rank a plurality of card-scan gradients relative to each other based on respective active-learning metrics;

based on the ranking, identify a low-information-entropy card-scan gradient and a high-information-entropy card-scan gradient; and

label characters corresponding to the low-information-entropy card-scan gradient and the high-information-entropy card-scan gradient indicate correct card characters by:

providing the low-information-entropy card-scan gradient to an automatic labeling system for labeling; and

providing the high-information-entropy card-scan gradient for visualization within a labeling interface for labeling.

19. The non-transitory computer readable medium of claim 18 , further comprising instructions that, when executed by the at least one processor, cause the computing device to update the card-scan machine learning model by modifying parameters of the card-scan machine learning model through retraining a stored version of the card-scan machine learning model to generate predictions of card characters based on the low-information-entropy card-scan gradient and the high-information-entropy card-scan gradient.

20. The non-transitory computer readable medium of claim 15 , further comprising instructions that, when executed by the at least one processor, cause the computing device to:

identify, via the client device, one or more of a logo, a hologram, or a chip on the character-bearing card by utilizing an object-detection algorithm of the card-scan machine learning model to process a set of card images;

determine a card number from the character-bearing card by utilizing a parser from the card-scan machine learning model to scan fields detected within the set of card images;

detect, via the client device, three-dimensional movement of the character-bearing card from the set of card images utilizing an optical-field detector of the card-scan machine learning model; and

based on the three-dimensional movement, the card number, and identification of one or more of the logo, the hologram, or the chip, cause the client device to utilize the card-scan machine learning model to verify that the character-bearing card associated with the set of card images is valid and a physical character-bearing card.