Computationally efficient document stamp detector
A digital image comparator includes an Faster R-CNN configured to generate a filtered set of local feature maps of an input image, a match head, and logic to preserve the local feature maps in an indexed data structure and to make the local feature maps retrievable by the match head via local feature map indexes.
1 . A digital image comparator comprising:
an Faster R-CNN configured to generate a filtered set of local feature maps of an input image;
a match head; and
logic to:
preserve the local feature maps in an indexed data structure and to make the local feature maps retrievable by the match head via local feature map indexes;
pair the local feature maps with the feature map of a registered stamp image; and
iterate through p×q unique pairs of the local feature maps and feature maps of registered stamp images, where p is a number of local feature maps, and q is a number of registered stamp images, and to process the unique pairs through the match head to generate a similarity score for each unique pair.
2 . The digital image comparator of claim 1 , the match head configured to generate similarity scores without engaging an image feature map encoder.
3 . The digital image comparator of claim 1 , the logic further configured to apply the similarity score to influence a document duplication process.
4 . A digital image comparison process comprising:
operating a Faster R-CNN on an input image to generate a filtered set of local feature maps of an input image;
preserving the local feature maps in an indexed data structure;
retrieving the local feature maps via local feature map indexes and applying the local feature maps to a match head, wherein the match head performs classification on a concatenation of one of the local feature maps and a second feature map for a second image, wherein the second feature map comprises a feature map for a registered stamp image; and
iterating through p×q unique pairs of the local feature maps and feature maps of registered stamp images, where p is a number of the local feature maps, and q is a number of registered stamp images, and processing the unique pairs through the match head to generate a similarity score for each unique pair.
5 . The process of claim 4 , the match head generating the similarity scores without engaging an image feature map encoder.
6 . The process of claim 4 , further comprising:
applying the similarity score to influence a document duplication process.