Granted Patent B2

US 12,488,562 · App. 17/536,400 · Granted Dec 2, 2025

Inventor: Amiteshwar Dayal Seth (Bangalore, IN)

Assignee: Accenture Global Solutions Limited

G06V10/761G06F16/532G06N20/00G06V10/42G06V10/44G06V10/7715G06V10/82

Abstract

Claims (74)

1 . A method for performing a visual search based on a query image, the method comprising:

receiving, by one or more processors, the query image;

applying, by the one or more processors, a fractal transform to the query image;

determining, by the one or more processors, an immediate search space based on the fractal transform of the query image, wherein the immediate search space corresponds to a subset of a plurality of images stored in a dataset;

applying, by the one or more processors, deep learning logic to the query image to produce a feature set for the query image,

wherein the feature set includes information associated with features of the query image;

evaluating, by the one or more processors, the feature set of the query image against feature sets of the subset of the plurality of images included in the immediate search space to determine a set of search results,

wherein the set of search results comprises one or more images from the subset of the plurality of images having features similar to the query image;

retrieving the features sets of the subset of the plurality of images included in the immediate search space from a memory based on the fractal transform of the query image,

wherein the feature set of the query image is determined using one or more kernels of a visual search engine, and

wherein a first portion of the feature set of the query image is obtained from the query image, a second portion of the feature set is obtained from a first derived set of images during training of the visual search engine, and a third portion of the feature set is obtained from a second derived set of images during training of the visual search engine; and

outputting, by the one or more processors, the set of search results.

2 . The method of claim 1 , wherein the deep learning logic comprises a first convolutional neural network and a second convolutional neural network,

wherein the first convolutional neural network is configured to identify high level features of an image,

wherein the second convolutional neural network is configured to identify low level features of the image,

wherein the high level features comprise categorical features of content of the image and the low level features comprise characteristics of the content of the image, and

wherein the high level features and low level features are included in the feature set.

3 . The method of claim 2 , wherein the high level features and the low level features are identified simultaneously.

4 . The method of claim 1 , wherein the set of search results are determined using a K-nearest neighbors algorithm.

5 . The method of claim 1 , wherein the fractal transform is performed using a partitioned iterative function system.

6 . The method of claim 5 , wherein a loss function is configured to receive the feature set as an input and output the set of search results.

7 . A non-transitory computer-readable storage medium storing instructions that, when executed by one or more processors, cause the one or more processors to perform operations to conduct a visual search based on a query image, the operations comprising:

receiving the query image;

applying a fractal transform to the query image;

determining an immediate search space based on the fractal transform of the query image, wherein the immediate search space corresponds to a subset of a plurality of images stored in a dataset;

applying deep learning logic to the query image to produce a feature set for the query image,

wherein the feature set includes information associated with features of the query image;

evaluating the feature set of the query image against feature sets of the subset of the plurality of images included in the immediate search space to determine a set of search results,

wherein the set of search results comprises one or more images from the subset of the plurality of images having features similar to the query image;

retrieving the features sets of the subset of the plurality of images included in the immediate search space from a memory based on the fractal transform of the query image,

wherein the feature set of the query image is determined using one or more kernels of a visual search engine, and

wherein a first portion of the feature set of the query image is obtained from the query image, a second portion of the feature set is obtained from a first derived set of images during training of the visual search engine, and a third portion of the feature set is obtained from a second derived set of images during training of the visual search engine; and

outputting the set of search results.

8 . The non-transitory computer-readable storage medium of claim 7 , wherein the deep learning logic comprises a first convolutional neural network and a second convolutional neural network,

wherein the first convolutional neural network is configured to identify high level features of an image,

wherein the second convolutional neural network is configured to identify low level features of the image,

wherein the high level features and the low level features are included in the feature set, and

wherein the high level features and the low level features are identified simultaneously.

9 . The non-transitory computer-readable storage medium of claim 8 , wherein the first convolutional network is a deep network and the second convolutional network is a shallow network, and wherein the deep network is configured to identify the high level features and the shallow network is configured to identify the low level features.

10 . The non-transitory computer-readable storage medium of claim 7 , the operations further comprising training a visual search engine, wherein the training comprises:

generating an image triplet based on a source image and additional images,

wherein the image triplet comprises the source image, a first derived set of images, and a second derived set of images,

wherein the first derived set of images are selected from the additional images based on a clustering algorithm, and

wherein the second derived set of images are selected from the additional images based on fractal sampling.

11 . The non-transitory computer-readable storage medium of claim 7 , wherein the deep learning logic comprises a loss function configured to receive the feature set as an input and output the set of search results.

12 . A system comprising:

a memory;

one or more processors communicatively coupled to the memory;

a dataset accessible to the one or more processors,

the dataset comprising a plurality of images;

a visual search engine executable by one or more processors, the visual search engine configured to:

apply a fractal transform to each image of the plurality of images;

apply deep learning logic to the plurality of images to produce a feature set for each image of the plurality of images,

wherein the feature set for each image of the plurality of images includes information associated with features of the image; and

store information associated with the fractal transform and the feature set in associated with a corresponding image such that each image of the dataset is associated with a fractal transform and a feature set to produce a set of searchable images;

determine, in response to receiving a query image, an immediate search space corresponding to a subset of the plurality of images based on the stored information and the fractal transform of the query image;

retrieving the features sets of the subset of the plurality of images included in the immediate search space from a memory based on the fractal transform of the query image,

wherein the feature set of the query image is determined using one or more kernels of the visual search engine, and

wherein a first portion of the feature set of the query image is obtained from the query image, a second portion of the feature set is obtained from a first derived set of images during training of the visual search engine, and a third portion of the feature set is obtained from a second derived set of images during training of the visual search engine; and

outputting the set of search results.

13 . The system of claim 12 , wherein the deep learning logic comprises a first convolutional neural network and a second convolutional neural network,

the first convolutional neural network is configured to identify high level features of an image and the second convolutional neural network is configured to identify low level features of the image,

wherein the high level features and the low level features are included in the feature set and are identified simultaneously.

14 . The system of claim 12 , wherein the visual search engine is configured to:

generate an image triplet based on each image of the plurality of images,

wherein the image triplet comprises a source image corresponding to a selected image of the plurality of images, a first derived set of images, and a second derived set of images,

wherein the first derived set of images is selected from the plurality of images based on a clustering algorithm and the second derived set of images is selected from the plurality of images based on fractal sampling.

15 . The system of claim 12 , wherein the plurality of images are logically grouped in the dataset into different categories according to a fractal-based index.

16 . The system of claim 12 , wherein the visual search engine is configured to:

receive the query image;

generate the fractal transform of the query image;

apply deep learning logic to the query image to produce a feature set for the query image;

compare the feature set for the query image to stored feature sets corresponding to images included in the immediate search space to identify a set of search results; and

output the set of search.

Continuity (2)

References Cited (18)