Method and electronic device with feature vector and feature map output

Patent metadata

Field	Value
Publication number	US-12541950-B2
Application number	US-202318310075-A
Country	US
Kind code	B2
Filing date	May 1, 2023
Priority date	Nov 1, 2022
Publication date	Feb 3, 2026
Grant date	Feb 3, 2026

Abstract

A processor-implemented method includes: determining distances between an input vector and center vectors comprised in a plurality of output nodes comprised in a trained codebook; and outputting a first feature vector of the input vector based on the distances between the center vectors and the input vector, wherein the trained codebook is trained by: determining a distance between a training input vector and the center vector for each of the output nodes; determining, among the plurality of output nodes, a best matched unit (BMU) in which a distance between the training input vector and the center vector of the BMU is minimized; and training the codebook by updating the center vector of the BMU, based on the distance between the training input vector and the center vector of the BMU.

First claim

What is claimed is:

1 . A processor-implemented method, the method comprising:
determining distances between an input vector and center vectors comprised in a plurality of output nodes comprised in a trained codebook; and outputting a first feature vector of the input vector based on the distances between the center vectors and the input vector, wherein the trained codebook is trained by:
determining a distance between a training input vector and the center vector for each of the output nodes;
determining, among the plurality of output nodes, a best matched unit (BMU) in which a distance between the training input vector and the center vector of the BMU is minimized; and
training the codebook by updating the center vector of the BMU and one or more other center vectors of the output nodes, based on the distance between the training input vector and the center vector of the BMU.

2 . The method of claim 1 , wherein the training of the codebook comprises:
determining distances between the plurality of output nodes and the BMU, using position coordinates of the plurality of output nodes; and training the codebook by updating the center vectors of the plurality of output nodes, based on the distances between the plurality of output nodes and the BMU.

3 . The method of claim 2 , wherein the updating of the center vectors of the plurality of output nodes comprises:
determining a plurality of weights based on the distances between the plurality of output nodes and the BMU; and updating the center vectors of the plurality of output nodes by respectively applying the weights to the center vectors.

4 . The method of claim 3 , wherein the determining of the weights comprises, in response to a first distance being greater than a second distance among the distances between the plurality of output nodes and the BMU, determining a first weight determined based on the first distance to be less than a second weight determined based on the second distance among the weights.

5 . The method of claim 1 , further comprising:
based on position coordinates of elements comprised in the first feature vector, determining an average of elements within a set range; and outputting a second feature vector, using the position coordinates of the elements and the average of the elements, wherein the position coordinates of the elements correspond to position coordinates of the plurality of output nodes.

6 . The method of claim 5 , further comprising outputting an output feature vector by inputting the second feature vector into a convolutional neural network.

7 . The method of claim 1 , further comprising generating the input vector by:
generating a plurality of segmented images from an input image, using a scan window having a set size; and generating the input vector, using the plurality of segmented images, wherein the outputting of the first feature vector of the input vector comprises outputting a first feature map of the input image, based on a distance between the center vector and the input vector for each of the plurality of segmented images.

8 . The method of claim 7 , further comprising:
based on position coordinates of channels comprised in the first feature map, determining an average of channels within a set range; and outputting a second feature vector, using the position coordinates of the channels and the average of the channels, wherein the position coordinates of the channels correspond to position coordinates of the plurality of output nodes.

9 . A non-transitory computer-readable storage medium storing instructions that, when executed by one or more processors, configure the one or more processors to perform the method of claim 1 .

10 . A processor-implemented method, the method comprising:
generating a plurality of segmented maps from an input map; determining distances between the plurality of segmentation maps and center vectors comprised in a plurality of output nodes comprised in a trained codebook; and outputting a first feature map of the input map, based on the distances between the plurality of segmentation maps and the center vectors, wherein the trained codebook is trained by:
determining distances between the center vectors and a training segmented map segmented from a training input map;
determining, among the plurality of output nodes, a best matched unit (BMU) in which a distance between the training segmented map and the center vector of the BMU is minimized; and
training the codebook by updating the center vector of the BMU, based on the distance between the training segmented map and the center vector of the BMU.

11 . The method of claim 10 , wherein the training of the codebook comprises:
determining distances between the plurality of output nodes and the BMU, using position coordinates of the plurality of output nodes; and training the codebook by updating the center vectors of the plurality of output nodes, based on the distances between the plurality of output nodes and the BMU.

12 . The method of claim 11 , wherein the updating of the center vectors of the plurality of output nodes comprises:
determining a plurality of weights based on the distances between the plurality of output nodes and the BMU; and updating the center vectors of the plurality of output nodes by respectively applying the weights to the center vectors.

13 . The method of claim 12 , wherein the determining of the weights comprises, in response to a first distance being greater than a second distance among the distances between the plurality of output nodes and the BMU, determining a first weight determined based on the first distance to be less than a second weight determined based on the second distance among the weights.

14 . The method of claim 10 , further comprising:
based on position coordinates of channels comprised in the first feature map, determining an average of channels within a set range; and outputting a second feature map, using the position coordinates of the channels and the average of the channels, wherein the position coordinates of the channels correspond to position coordinates of the plurality of output nodes.

15 . An electronic device comprising:
one or more processors configured to:
determine distances between an input vector and center vectors comprised in a plurality of output nodes comprised in a trained codebook; and
output a first feature vector of the input vector, based on the distances between the center vectors and the input vector, wherein the trained codebook is trained by:
determining a distance between a training input vector and the center vector for each of the output nodes;
determining, among the plurality of output nodes, a best matched unit (BMU) in which a distance between the training input vector and the center vector of the BMU is minimized; and
training the codebook by updating the center vector of the BMU and one or more other center vectors of the output nodes, based on the distance between the training input vector and the center vector of the BMU.

16 . The electronic device of claim 15 , wherein the training of the codebook comprises:
determining distances between the plurality of output nodes and the BMU, using position coordinates of the plurality of output nodes; and training the codebook by updating the center vectors of the plurality of output nodes, based on the distances between the plurality of output nodes and the BMU.

17 . The electronic device of claim 15 , wherein the one or more processors are configured to:
based on position coordinate

Assignees

• Samsung Electronics Co Ltd

Inventors

• Kim Hyunsoo
• Kim Seijoon
• Mun Minyoung
• Yoon Seungkeun

Classifications

• G06V10/82

  using neural networks · CPC title

• G06V10/774

  Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title

• G06V10/761

  Proximity, similarity or dissimilarity measures · CPC title

• G06V10/50 Primary

  by performing operations within image blocks; by using histograms, e.g. histogram of oriented gradients [HoG]; by summing image-intensity values; Projection analysis · CPC title

• G06V10/7715 Primary

  Feature extraction, e.g. by transforming the feature space, e.g. multi-dimensional scaling [MDS]; Mappings, e.g. subspace methods · CPC title