Image recognition method and apparatus based on systolic array, and medium

Patent metadata

Field	Value
Publication number	US-11830244-B2
Application number	US-202118034013-A
Country	US
Kind code	B2
Filing date	Apr 26, 2021
Priority date	Oct 27, 2020
Publication date	Nov 28, 2023
Grant date	Nov 28, 2023

Abstract

An image recognition method and apparatus based on a systolic array, and a medium are disclosed. The method includes: converting obtained image feature information into a one-dimensional feature vector; converting an obtained weight matrix into a one-dimensional weight vector, and allocating a corresponding weight group to each node in a trained three-dimensional systolic array model; performing multiply-accumulate of the feature vector and a weight value on the one-dimensional feature vector in parallel by using the three-dimensional systolic array model, to obtain a feature value corresponding to each node, the feature values with different values reflecting article categories contained in an image; and determining an article category contained in the image according to the feature value corresponding to each node and a pre-established corresponding relationship between the feature value and the article category.

First claim

What is claimed is:

1. An image recognition method based on a systolic array, comprising:
converting obtained image feature information into a one-dimensional feature vector;
converting an obtained weight matrix into a one-dimensional weight vector, and allocating a corresponding weight group to each node in a trained three-dimensional systolic array model;
performing multiply-accumulate of a feature vector and a weight value on the one-dimensional feature vector in parallel by using the three-dimensional systolic array model, to obtain a feature value corresponding to each node; and
determining an article category contained in the image according to the feature value corresponding to each node and a pre-established corresponding relationship between the feature value and the article category.

2. The image recognition method based on the systolic array according to claim 1 , wherein the performing multiply-accumulate of a feature vector and a weight value on the one-dimensional feature vector in parallel by using the three-dimensional systolic array model, to obtain a feature value corresponding to each node comprises:
performing multiply-accumulate on the one-dimensional feature vector and a current feature value of a first node in the three-dimensional systolic array model, and adopting an obtained multiply-accumulate value as the feature value of the first node;
transmitting, according to a node transmission direction of the three-dimensional systolic array model, the one-dimensional feature vector respectively to at least three second nodes that are directly connected to the first node; and
performing multiply-accumulate on the one-dimensional feature vector and a current feature value of each of the second nodes separately, adopting an obtained multiply-accumulate value as the feature value separately corresponding to each of the second nodes, and adopting each of the second nodes as the first node, and returning to the step of performing multiply-accumulate on the one-dimensional feature vector and a current feature value of a first node in the three-dimensional systolic array model, and adopting an obtained multiply-accumulate value as the feature value of the first node, until calculation of feature values of all nodes in the three-dimensional systolic array model is completed.

3. The image recognition method based on the systolic array according to claim 2 , wherein after the performing multiply-accumulate of a feature vector and a weight value on the one-dimensional feature vector in parallel by using the three-dimensional systolic array model, to obtain a feature value corresponding to each node, the method further comprises:
judging whether a number of currently obtained feature values reaches an output number corresponding to the image feature information;
in response to the number of currently obtained feature values not reaching the output number corresponding to the image feature information, adopting a next weight value adjacent to a current weight value in the weight group of each node in the three-dimensional systolic array model as a latest current weight value of each node, and returning to the step of performing multiply-accumulate on the one-dimensional feature vector and a current feature value of a first node in the three-dimensional systolic array model, and adopting an obtained multiply-accumulate value as the feature value of the first node; and
wherein the determining an article category contained in the image according to the feature value corresponding to each node and a pre-established corresponding relationship between the feature value and the article category is executed in response to the number of currently obtained feature values reaching the output number corresponding to the image feature information.

4. The image recognition method based on the systolic array according to claim 1 , wherein the determining an article category contained in the image according to the feature value corresponding to each node and a pre-established corresponding relationship between the feature value and the article category comprises:
calculating a cumulative sum value of feature values of all nodes according to the feature value corresponding to each node;
calculating a ratio of the feature value corresponding to each node to the cumulative sum value, and selecting a node with the ratio satisfying a preset condition as a target node; and
querying the pre-established corresponding relationship between the feature value and the article category, to determine the article category that matches the feature value of the target node.

5. The image recognition method based on the systolic array according to claim 4 , wherein the selecting a node with the ratio satisfying a preset condition as a target node comprises:
selecting a node with the largest ratio as the target node.

6. The image recognition method based on the systolic array according to claim 1 , wherein before the allocating a corresponding weight group to each node in a trained three-dimensional systolic array model, the method further comprises:
adjusting a number of channels of the three-dimensional systolic array model according to an output number corresponding to the image feature information.

7. The image recognition method based on the systolic array according to claim 2 , wherein before the allocating a corresponding weight group to each node in a trained three-dimensional systolic array model, the method further comprises:
adjusting a number of channels of the three-dimensional systolic array model according to an output number corresponding to the image feature information.

8. The image recognition method based on the systolic array according to claim 3 , wherein before the allocating a corresponding weight group to each node in a trained three-dimensional systolic array model, the method further comprises:
adjusting a number of channels of the three-dimensional systolic array model according to an output number corresponding to the image feature information.

9. An image recognition apparatus based on a systolic array, comprising:
a memory storing a computer program; and
a processor, configured to execute the computer program, and upon execution of the computer program, is configured to:
convert obtained image feature information into a one-dimensional feature vector;
convert an obtained weight matrix into a one-dimensional weight vector, and allocate a corresponding weight group to each node in a trained three-dimensional systolic array model;
perform multiply-accumulate of a feature vector and a weight value on the one-dimensional feature vector in parallel by using the three-dimensional systolic array model, to obtain a feature value corresponding to each node; and
determine an article category contained in the image according to the feature value corresponding to each node and a pre-established corresponding relationship between the feature value and the article category.

10. The image recognition apparatus according to claim 9 , wherein in order to perform the multiply-accumulate of a feature vector and a weight value on the one-dimensional feature vector in parallel by using the three-dimensional systolic array model, to obtain a feature value corresponding to each node, the processor, upon execution of the computer program, is configured to:
perform multiply-accumulate on the one-dimensional feature vector and a current feature value of a first node in the three-dimensional systolic array model, and adopt an obtained multiply-accumulate value as the feature value of the first node;
transmit, according to a node transmission direction of the three-dimensional systolic array model, the one-dimensional fea

Assignees

• Inspur Suzhou Intelligent Technology Co Ltd

Inventors

• Dong Gang
• Zhao Yaqian
• Li Rengang
• Yang Hongbin
• Liu Haiwei
• Jiang Dongdong

Classifications

• G06V10/955 Primary

  using specific electronic processors · CPC title

• G06V10/751

  Comparing pixel values or logical combinations thereof, or feature values having positional relevance, e.g. template matching · CPC title

• G06V10/764 Primary

  using classification, e.g. of video objects · CPC title

• G06V10/7715

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

• G06F18/21 Primary

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