Method, device, and computer program product for image recognition

Patent metadata

Field	Value
Publication number	US-12561970-B2
Application number	US-202117405241-A
Country	US
Kind code	B2
Filing date	Aug 18, 2021
Priority date	Jul 23, 2021
Publication date	Feb 24, 2026
Grant date	Feb 24, 2026

Abstract

Embodiments of the present disclosure relate to a method, a device, and a computer program product for image recognition. In some embodiments, characterization information for a first reference image in a reference image set is generated in an image recognition engine by using a Gaussian mixture model. First reference label information for the first reference image is generated based on the characterization information for the first reference image, the first reference label information being associated with a category of a first object in the first reference image. The image recognition engine is updated by determining the accuracy of the first reference label information for the first reference image. In this way, good characterization of images and generation of reference label information for the images can be achieved, thus both improving the robustness of the generated reference label information and significantly improving the accuracy of image recognition.

First claim

What is claimed is:

1 . A method for image recognition, including:
in a first training stage in a sequence of multiple training stages of a processor-based machine learning system: (i) generating, in an image recognition engine implemented in the processor-based machine learning system, characterization information for a first reference image in a reference image set by using a Gaussian mixture model, wherein a plurality of weights are generated for the Gaussian mixture model as respective orthogonal bases extracted from at least one layer of at least one neural network of the image recognition engine; (ii) generating first reference label information in the processor-based machine learning system for the first reference image based on the characterization information for the first reference image, the first reference label information being associated with a category of a first object in the first reference image; and (iii) updating the image recognition engine implemented in the processor-based machine learning system by determining the accuracy of the first reference label information for the first reference image, wherein updating the image recognition engine comprises performing an iterative optimization of the orthogonal bases in an expectation maximization feedback loop implemented by a processor of the processor-based machine learning system, the expectation maximization feedback loop comprising an expectation execution path in which one or more of the orthogonal bases are adjusted by the processor and a maximization execution path in which one or more parameters of a classifier are adjusted by the processor based on the one or more adjusted orthogonal bases, an output of the expectation execution path being coupled to an input of the maximization execution path and an output of the maximization execution path being coupled to an input of the expectation execution path, the iterative optimization repeatedly executing the expectation execution path and the maximization execution path over a plurality of iterations until a designated accuracy is reached, to obtain an updated image recognition engine; and in a second training stage following the first training stage in the sequence of multiple training stages of the processor-based machine learning system: (i) retraining the updated image recognition engine utilizing one or more additional reference images having respective additional reference label information associated therewith; wherein one or more additional images are classified utilizing the classifier in the retrained image recognition engine, subsequent to execution of the first and second training stages in the sequence of multiple training stages of the processor-based machine learning system.

2 . The method according to claim 1 , further including:
generating, in the second training stage and based on a second reference image without label information in the reference image set and by means of the updated image recognition engine, second reference label information for the second reference image, the second reference label information being associated with a category of an object in the second reference image; and using the second reference image and the second reference label information for the second reference image to further update the image recognition engine in the second training stage.

3 . The method according to claim 1 , wherein generating the characterization information for the first reference image includes:
generating weight information for the Gaussian mixture model based on the first reference image and initial expectation information for the Gaussian mixture model; and generating the characterization information for the first reference image based on the initial expectation information and the weight information.

4 . The method according to claim 3 , wherein generating the weight information for the Gaussian mixture model includes:
generating the weight information for the Gaussian mixture model using a kernel function based on the first reference image and the initial expectation information for the Gaussian mixture model.

5 . The method according to claim 3 , wherein the initial expectation information is a multi-dimensional vector, and vectors of at least two dimensions in the multi-dimensional vector are orthogonal.

6 . The method according to claim 3 , further including:
obtaining the initial expectation information for the Gaussian mixture model based on a third reference image with label information in the reference image set and third pre-existing label information for the third reference image, the third pre-existing label information being associated with a category of a third object in the third reference image.

7 . The method according to claim 3 , further including:
generating a reconstructed image based on the initial expectation information for the Gaussian mixture model; and updating the initial expectation information based on the reconstructed image.

8 . A device for image recognition, including:
a processor, and a memory storing computer-executable instructions, wherein the computer-executable instructions, when executed by the processor, cause the device to perform actions including: in a first training stage in a sequence of multiple training stages of a processor-based machine learning system: (i) generating, in an image recognition engine implemented in the processor-based machine learning system, characterization information for a first reference image in a reference image set by using a Gaussian mixture model, wherein a plurality of weights are generated for the Gaussian mixture model as respective orthogonal bases extracted from at least one layer of at least one neural network of the image recognition engine; (ii) generating first reference label information in the processor-based machine learning system for the first reference image based on the characterization information for the first reference image, the first reference label information being associated with a category of a first object in the first reference image; and (iii) updating the image recognition engine implemented in the processor-based machine learning system by determining the accuracy of the first reference label information for the first reference image, wherein updating the image recognition engine comprises performing an iterative optimization of the orthogonal bases in an expectation maximization feedback loop implemented by a processor of the processor-based machine learning system, the expectation maximization feedback loop comprising an expectation execution path in which one or more of the orthogonal bases are adjusted by the processor and a maximization execution path in which one or more parameters of a classifier are adjusted by the processor based on the one or more adjusted orthogonal bases, an output of the expectation execution path being coupled to an input of the maximization execution path and an output of the maximization execution path being coupled to an input of the expectation execution path, the iterative optimization repeatedly executing the expectation execution path and the maximization execution path over a plurality of iterations until a designated accuracy is reached, to obtain an updated image recognition engine; and in a second training stage following the first training stage in the sequence of multiple training stages of the processor-based machine learning system: (i) retraining the updated image recognition engine utilizing one or more additional reference images having respective additional reference label information associated therewith; wherein one or more additional images are classified utilizing the classifier in the retrained image recognition engine,

Assignees

• Emc Ip Holding Co Llc

Inventors

• Wang Zijia
• Ni Jiacheng
• Jia Zhen
• Yang Wenbin

Classifications

• G06V20/80

  Recognising image objects characterised by unique random patterns · CPC title

• G06N7/01

  Probabilistic graphical models, e.g. probabilistic networks · CPC title

• G06F18/214

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

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/084

  Backpropagation, e.g. using gradient descent · CPC title