US 10,019,654 B1Granted

Image object recognition

AssigneeAccenture Global Solutions Limited

InventorAndrea Pisoni

Life of this patent

Estimating the term…

Loading patent…

Abstract

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for recognizing object sub-types in images. One of the methods includes receiving training data; selecting training data for an image; determining whether to randomly permute a value of a property of the selected image; providing, to a deep neural network, the particular training data or the randomly permuted particular training data; receiving, from the deep neural network, output data indicating a predicted label for an object sub-type for an object depicted in the selected image, and a confidence score that represents a likelihood that the object has the object sub-type; updating one or more weights in the deep neural network using an expected output value, the predicted label, and the confidence score; and providing the deep neural network to a mobile device for use detecting whether one or more images depict objects having the particular object sub-type.

Claims20

Dependent claims

A computer-implemented method comprising:

receiving training data representing a plurality of images at a deep neural network training system,
the plurality of images depicting objects having an object sub-type, and
the deep neural network training system to train a deep neural network to recognize the object sub-type for one of the objects depicted in the plurality of images,
wherein multiple object sub-types for which corresponding objects are depicted in the plurality of images each correspond to the same object type;
for a plurality of iterations:
selecting particular training data for an image from the plurality of images;
determining whether to randomly permute a value of a property of the selected image that is represented by the particular training data;
providing, to the deep neural network, the particular training data for the selected image or the randomly permuted particular training data for the selected image;
receiving, from the deep neural network, output data indicating:
a predicted label for a particular object sub-type for an object depicted in the selected image, and
a confidence score that represents a likelihood that the object depicted in the selected image has the particular object sub-type;
and
updating one or more weights in the deep neural network using:
an expected output value that indicates an actual object sub-type label for the object depicted in the selected image,
the predicted label for the particular object sub-type, and
the confidence score that represents the likelihood that the object depicted in the selected image has the particular object sub-type;
and
providing the deep neural network to a mobile device for use detecting whether one or more images depict objects having the particular object sub-type.

15 more elements

A non-transitory computer storage medium encoded with instructions that, when executed by one or more computers, cause the one or more computers to perform operations comprising:

receiving data for an image that depicts a plurality of objects;
determining whether to create multiple overlapping portions of data for the image,
the multiple overlapping portions each including a proper subset of data for the image;
in response to determining to create the multiple overlapping portions of data for the image:
creating the multiple overlapping portions of data for the image using the received data for the image;
providing, to a deep neural network, data for each of the multiple overlapping portions separately to cause the deep neural network to predict object sub-type labels for each of the plurality of objects;
receiving, from the deep neural network for some of the objects in the plurality of objects:
output data that represents a sub-type label that indicates a likely object sub-type for the object, and
an object location that indicates a location of the object in the image;
and
generating, using the output data for some of the objects in the plurality of objects, a representation of the image that indicates, for some of the objects in the plurality of objects, a location of the object in the image and the sub-type label that indicates the likely object sub-type for the object.

8 more elements

A deep neural network training system comprising one or more computers and one or more storage devices on which are stored instructions that are operable, when executed by the one or more computers, to cause the one or more computers to perform operations comprising:

receiving training data representing a plurality of images,
the plurality of images depicting objects having an object sub-type, and
the deep neural network training system training a deep neural network to recognize the object sub-type for one of the objects depicted in the plurality of images,
wherein multiple object sub-types for which corresponding objects are depicted in the plurality of images each correspond to the same object type;
for a plurality of iterations:
selecting particular training data for an image from the plurality of images;
determining whether to randomly permute a property of the selected image that is represented by the particular training data;
providing, to the deep neural network, the particular training data for the selected image or the randomly permuted particular training data for the selected image;
receiving, from the deep neural network, output data indicating:
a predicted label for a particular object sub-type for an object depicted in the selected image,
location information defining a region of the selected image in which the object is likely depicted, and
a confidence score that represents a likelihood that the object exists in the region of the selected image defined by the location information;
and
updating one or more weights in the deep neural network using:
an expected output value that indicates an actual object sub-type label for the object depicted in the selected image,
the predicted label for the particular object sub-type, and
the confidence score that represents the likelihood that the object exists in the region of the selected image defined by the location information;
and
providing the deep neural network to a mobile device for use detecting whether one or more images depict objects having the particular object sub-type.

16 more elements

Dependencies are inferred from the claim text.

Facts

Application: 15/692,180Filed 31 Aug 2017 · us provisional application
Priority: 28 Jun 2017
Granted: 10 Jul 2018
Term ends: Not estimatedThe source records don’t support an estimate.

Classification17 codes

G06N 3/084
Learning methods: backpropagation, e.g. using gradient descent
G06F 18/214
Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation: generating training…
G06F 18/217
Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation: validation; Performance…
G06F 18/24
Analysing: classification techniques