Item recognition processing over time
Time-stamped images are captured for unknown items within a store. Portions of the images are cropped to restrict the cropped images to attributes associated with the unknown items. A set of time ordered cropped images are proved to a trained machine-learning item recognition algorithm, which produces an output of confidence values that map to specific items of the store. When the confidence value meets or exceeds a predefined threshold, the specific item associated with that confidence value is used to identify the unknown item. The known item is assigned to an appropriate shopping cart of a shopper based on tracking from the images of the shopper while in the store.
1. A method, comprising:
cropping images into cropped images, each image comprises at least a portion of an unknown item and modifying the cropped images by removing pixels from each of the cropped images that are known to be associated with background objects and that are known to be associated with persons leaving only remaining pixels in each of the cropped images that just comprise item attributes for the unknown item within the corresponding cropped image;
inputting the cropped images into a trained machine-learning item identifier;
receiving from the trained machine-learning item identifier a list that includes confidences values, each confidence value associated with an item identifier;
selecting a particular item identifier for the unknown item based on a highest confidence value from the confidence values of the list; and
providing the particular item identifier as a known item for the unknown item to a transaction manager.
2. The method of claim 1 , wherein cropping further includes maintaining with each cropped image a frame number and a time stamp.
3. The method of claim 2 , wherein maintaining further includes generating a unknown item tracking identifier for the unknown item and maintaining the unknown item tracking identifier with each cropped image.
4. The method of claim 3 , wherein generating further includes organizing the cropped images as a time-ordered set of cropped images.
5. The method of claim 4 , wherein inputting further includes providing the time-ordered set of cropped images as an input to the trained machine-learning item identifier.
6. The method of claim 1 , wherein receiving further includes sorting the list from the highest confidence value to a lowest confidence value.
7. The method of claim 6 , wherein selecting further includes obtaining a first entry from the list and obtaining from the first entry the highest confidence value and the particular item identifier.
8. The method of claim 1 , wherein selecting further includes comparing the highest confidence value against a threshold value and selecting the highest confidence value when the highest confidence value is equal to or exceeds the threshold value.
9. The method of claim 8 , wherein comparing further includes selecting the threshold value from a plurality of available threshold values based on the particular item identifier.