LATENT EMBEDDINGS FOR WORD IMAGES AND THEIR SEMANTICS
First Claim
Patent Images
1. A semantic comparison method, comprising:
- providing training word images labeled with concepts;
with the training word images and their labels, learning a first embedding function for embedding word images in a semantic subspace into which the concepts are embedded with a second embedding function;
receiving a query comprising at least one test word image or at least one concept;
where the query comprises at least one test word image, generating a representation of each of the at least one test word image, comprising embedding the test word image in the semantic subspace with the first embedding function;
where the query comprises at least one concept, providing a representation of the at least one concept generated by embedding each of the at least one concept the embedding function;
computing a comparison between;
a) at least one of the test word image representations, andb) at least one of;
at least one of the concept representations, andanother of test word image representations; and
outputting information based on the comparison.
6 Assignments
0 Petitions
Accused Products
Abstract
A system and method enable semantic comparisons to be made between word images and concepts. Training word images and their concept labels are used to learn parameters of a neural network for embedding word images and concepts in a semantic subspace in which comparisons can be made between word images and concepts without the need for transcribing the text content of the word image. The training of the neural network aims to minimize a ranking loss over the training set where non relevant concepts for an image which are ranked more highly than relevant ones penalize the ranking loss.
107 Citations
20 Claims
-
1. A semantic comparison method, comprising:
-
providing training word images labeled with concepts; with the training word images and their labels, learning a first embedding function for embedding word images in a semantic subspace into which the concepts are embedded with a second embedding function; receiving a query comprising at least one test word image or at least one concept; where the query comprises at least one test word image, generating a representation of each of the at least one test word image, comprising embedding the test word image in the semantic subspace with the first embedding function; where the query comprises at least one concept, providing a representation of the at least one concept generated by embedding each of the at least one concept the embedding function; computing a comparison between; a) at least one of the test word image representations, and b) at least one of; at least one of the concept representations, and another of test word image representations; and outputting information based on the comparison. - View Dependent Claims (2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17)
-
-
18. A semantic comparison system, comprising:
-
memory which stores a neural network having parameters which have been trained with training word images labeled with concepts from a set of concepts, each the concepts corresponding to a set of entity names, the training word images each being an image of one of the set of entity names for the concept with which it is labeled, the neural network having been learned to embed the training word images and the concepts into a common semantic space with a ranking loss objective function which favors the concepts that are relevant to a word image being ranked, by the neural network, ahead of those that are not relevant; a comparison component for computing a compatibility between a word image and a concept which have both been embedded in the common semantic space using the trained neural network; an output component which outputs information based on the comparison; and a processor in communication with the memory which implements the comparison component and output component. - View Dependent Claims (19)
-
-
20. A semantic comparison method, comprising:
-
providing a neural network having parameters which have been learned with training word images labeled with concepts from a set of concepts, the neural network having been learned to embed the training word images and the concepts into a common semantic space with a ranking loss objective function which favors the concepts that are relevant to a training word image being ranked, by the neural network, ahead of those that are not relevant; computing a compatibility between a word image and a concept which have both been embedded in the common semantic space using the trained neural network; and outputting information based on the compatibility computation.
-
Specification