Method, apparatus and device for extracting image features, and storage medium

Patent metadata

Field	Value
Publication number	US-12555347-B2
Application number	US-202118246299-A
Country	US
Kind code	B2
Filing date	Sep 28, 2021
Priority date	Jan 7, 2021
Publication date	Feb 17, 2026
Grant date	Feb 17, 2026

Abstract

Provided are a method, apparatus and device for extracting image features, and a storage medium. The method includes: obtaining parameters to be quantized of a network layer in a neural network model; determining whether values of the parameters to be quantized are all positive numbers; when the values of the parameters to be quantized are all positive numbers, executing, based on an asymmetric linear quantization logic, a quantization operation on the parameters to be quantized; when the values of the parameters to be quantized are not all positive numbers, executing, based on a symmetric linear quantization logic, a quantization operation on the parameters to be quantized; and extracting features of an input image by using the neural network model for which the quantization operation has been executed.

First claim

What is claimed is:

1 . A method for extracting image features, comprising:
obtaining parameters to be quantized of a network layer in a neural network model; determining whether values of the parameters to be quantized are all positive numbers; when the values of the parameters to be quantized are all positive numbers, executing, based on an asymmetric linear quantization logic, a quantization operation on the parameters to be quantized; when the values of the parameters to be quantized are not all positive numbers, executing, based on a symmetric linear quantization logic, a quantization operation on the parameters to be quantized; and extracting features of an input image by using the neural network model for which the quantization operation has been executed; wherein the method for extracting image features further comprises: correcting counted values of nodes having a data point-by-point addition or concat operation that execute operations comprising shortcut and concat, to make quantization coefficients of respective converged branches consistent; wherein the correcting counted values of nodes having a data point-by-point addition or concat operation that execute operations comprising shortcut and concat, to make quantization coefficients of respective converged branches consistent comprises: determining the nodes having the data point-by-point addition or concat operation that execute the operations comprising shortcut and concat as nodes to be corrected; selecting a maximum value in counted values of all precursor nodes corresponding to a current node and setting the selected maximum value as counted values of all the precursor nodes, wherein the current node is any one of all the nodes to be corrected; executing an exclusive OR operation on identity values of all the precursor nodes corresponding to the current node to obtain an operation value; when the operation value is an asymmetric identity value, updating output identity values of all the precursor nodes to the operation value, and updating input identity values of nodes, which take outputs of the precursor nodes as inputs, as the operation value; and when the operation value is a symmetric identity value, updating both an input identity value and an output identity value of the current node to the operation value, and updating input identity values of nodes, which take an output of the current node as an input, to the operation value.

2 . The method for extracting image features according to claim 1 , wherein the executing a quantization operation on the parameters to be quantized comprises:
for initial weight quantized values of the network layer, selecting each value in a first preset value interval as a rounding point; determining whether an absolute value of a decimal part of each initial weight quantized value is greater than the rounding point; when the absolute value of the decimal part of the initial weight quantized value is greater than the rounding point, rounding up the initial weight quantized value to obtain a target weight quantized value; otherwise, rounding down the initial weight quantized value to obtain the target weight quantized value; convolving the target weight quantized values and input features of the network layer to obtain convolution values, and computing mean square errors based on the convolution values and an original convolution value; and setting a target weight quantized value having a smallest mean square error among the convolution values as a final weight quantized value.

3 . The method for extracting image features according to claim 2 , further comprising:
quantizing initial bias values according to the weight quantization coefficients and quantization coefficients of the input features to obtain quantized bias values.

4 . The method for extracting image features according to claim 3 , wherein the executing a quantization operation on the parameters to be quantized comprises:
computing a scaling coefficient (SC) according to an equation

SC
=

s
0

s
w

*

s
i

,

wherein s o represents an activation quantization coefficient, s w represents the weight quantization coefficient, s i represents the quantization coefficient of the input features, and the activation quantization coefficient is computed according to a counted value corresponding to a node and an activation bandwidth value set by the counted value;
computing a displacement value n according to an equation SC=2 −n ;
executing a convolution operation on the weight quantized values and quantized values of the input features to obtain an operation result, and using sum values of the operation result and the quantized bias values as initial convolution results; and
executing displacement processing on the initial convolution results according to the displacement value to obtain activation quantized values.

5 . The method for extracting image features according to claim 4 , wherein the initial convolution results Y are computed according to a following equation:
Y =conv( W,I )+ B q , wherein W represents the weight quantized value, I represents the quantized value of the input features, and B q represents a quantized bias value.

6 . The method for extracting image features according to claim 2 , wherein the initial weight quantized values are computed according to weight values and weight quantization coefficients corresponding to the weight values, and the weight quantization coefficients are computed according to counted values corresponding to weights and weight bandwidth values set by the counted values.

7 . The method for extracting image features according to claim 6 , wherein the initial weight quantized value is

Weight

s
w

,

wherein Weight represents a weight value, and s w represents the weight quantization coefficient.

8 . The method for extracting image features according to claim 6 , wherein the weight quantization coefficient is s w =ceil(log 2 s), wherein

s
=

R
max

2
k

,

R max represents the counted value corresponding to the weight, when symmetric quantization is used, k=7, and when asymmetric quantization is used, k=8.

9 . The method for extracting image features according to claim 1 , wherein before the extracting features of an input image by using the neural network model for which the quantization operation has been executed, the method further comprises:
normalizing the input image.

10 . The method for extracting image features according to claim 1 , wherein the parameters to be quantized of the network layer are parameters in the

Assignees

• Inspur Suzhou Intelligent Technology Co Ltd

Inventors

• Zhao Jian
• Dong Gang
• Shi Hongzhi
• Cao Qichun
• Cui Xingchen

Classifications

• G06V10/82 Primary

  using neural networks · CPC title

• G06V10/44 Primary

  Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components · CPC title

• G06V10/28

  Quantising the image, e.g. histogram thresholding for discrimination between background and foreground patterns · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06N3/08

  Learning methods · CPC title