Method and image processing entity for applying a convolutional neural network to an image

Patent metadata

Field	Value
Publication number	US-10832076-B2
Application number	US-201816208587-A
Country	US
Kind code	B2
Filing date	Dec 4, 2018
Priority date	Dec 14, 2017
Publication date	Nov 10, 2020
Grant date	Nov 10, 2020

Abstract

A method and an image processing entity for applying a convolutional neural network to an image are disclosed. The image processing entity processes the image while using the convolutional kernel to render a feature map, whereby a second feature map size of the feature map is greater than a first feature map size of the feature maps with which the feature kernel was trained. Furthermore, the image processing entity repeatedly applies the feature kernel to the feature map in a stepwise manner, wherein the feature kernel was trained to identify the feature based on the feature maps of the first feature maps, wherein the feature kernel has the first feature map size.

First claim

The invention claimed is:

1. A method for applying a convolutional neural network to an image, wherein the convolutional neural network comprises a convolutional kernel for convolving with images of a first image size to render feature maps and a feature kernel trained to identify a feature in the images based on the feature maps, wherein the first image size is less than a second image size of the image to which the convolutional neural network is applied, wherein the method comprises:
processing the image while using the convolutional kernel to render a feature map, whereby a second feature map size of the feature map is greater than a first feature map size of the feature maps with which the feature kernel was trained,
selecting the feature kernel among a plurality of feature kernels based on a position of the feature kernel relatively the feature map, which feature kernel is trained for said position, wherein the plurality of feature kernels includes a respective feature kernel that have been trained at a respective one of nine positions within the feature map, and
repeatedly applying the feature kernel to the feature map in a stepwise manner, referring to displacement of consecutive applications of the feature kernel to the feature map, wherein the feature kernel was trained to identify the feature based on the feature maps of the first feature map size, wherein the feature kernel has the first feature map size, wherein the feature maps were obtained by convolving the convolutional kernel over images having the first image size, which causes, at least due to the convolution, the feature map to have the second feature map size, wherein the stepwise manner is represented by a step size that is greater than half of the first feature map size.

2. The method according to claim 1 , wherein the position is one of four different corner positions, four different edge positions and an internal position.

3. The method according to claim 1 , wherein at least two consecutive applications of the feature kernel are applied such that the step size is less than of equal to the first feature map size.

4. The method according to claim 1 , wherein the method is performed by an image processing entity.

5. An image processing entity configured for performing the method according to claim 1 .

6. A non-transitory computer storage medium that has computer readable code units stored therein that when executed on an image processing circuitry causes the image processing circuitry to perform the method according to claim 1 .

Assignees

• Axis Ab

Inventors

• Danielsson Niclas
• Molin Simon
• Skans Markus

Classifications

• G06V10/24 Primary

  Aligning, centring, orientation detection or correction of the image · CPC title

• G06F18/21

  Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06V10/454

  Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN] · CPC title

• G06N3/08

  Learning methods · CPC title