Target re-identification method, network training method thereof, and related device

Patent metadata

Field	Value
Publication number	US-12482237-B2
Application number	US-202218146450-A
Country	US
Kind code	B2
Filing date	Dec 26, 2022
Priority date	Jun 29, 2020
Publication date	Nov 25, 2025
Grant date	Nov 25, 2025

Abstract

A device for transferring micro-components and a method of manufacturing the device are provided. The device includes a substrate, a metal wire, and a plurality of silicon electrodes. The metal wire is formed on a flat surface of the substrate and includes a plurality of electrode driving units. The silicon electrodes are formed on a side of the metal wire opposing to the substrate. Each silicon electrode corresponds to each electrode driving unit and is driven by the electrode driving unit to pick or release each micro-component. According to the present application, the device may electrostatically adsorb a massive amount of micro-components to achieve transferring the massive amount of the micro-components, dramatically improving a transfer efficiency.

First claim

What is claimed is:

1 . A method for training a target re-identification network, comprising:
obtaining a training image set, wherein the training image set comprises a plurality of training images; for each of the plurality of training images in the training image set: identifying a training image based on the target re-identification network to obtain an identification result of the training image; wherein the target re-identification network comprises a plurality of branches, the identification result of the training image comprises a feature information output by each of the plurality of branches and a classification result corresponding to the feature information; the feature information output by one of the plurality of branches comprises n local feature information, the n being an integer greater than 3; the n local feature information correspond to different regions of the training image; and adjusting parameters of each of the plurality of branches of the target re-identification network based on the identification result.

2 . The method according to claim 1 , wherein the number of the plurality of branches is n, the feature information output by a first branch does not comprise the local feature information, the feature information output by an i-th branch comprises i local feature information local feature information; regions corresponding to the i local feature information local feature information of the i-th branch cover the plurality of training images; i is greater than 1 and less than or equal to n; and/or,
the feature information output by each of the plurality of branches comprises a global feature information.

3 . The method according to claim 1 , wherein the adjusting parameters of each of the plurality of branches of the target re-identification network based on the identification result comprises:
calculating a loss of each of the plurality of branches of the target re-identification network based on the identification result of the training image; and adjusting the parameters of each of the plurality of branches of the target re-identification network based on the loss.

4 . The method of claim 3 , wherein each of the plurality of branches comprises a convolutional layer, a sampling layer, and a feature embedding layer; the identifying the training image based on the target re-identification network to obtain an identification result of the training image comprises:
for each of the plurality of branches of the target re-identification network: processing the training image based on the convolutional layer of a branch to obtain a first feature information of the training image; processing the first feature information based on the sampling layer of the branch to obtain a second feature information of the training image; and processing the second feature information based on the feature embedding layer of the branch to obtain a third feature information of the training image, the third feature information being configured as the feature information output by the branch; wherein the calculating a loss of each of the plurality of branches of the target re-identification network based on the identification result of the training image comprises: calculating the feature information output by the branch based on a first loss function to obtain a first loss of the branch of the target re-identification network.

5 . The method of claim 3 , wherein each of the plurality of branches comprises a convolutional layer, and a sampling layer; the identifying the training image based on the target re-identification network to obtain an identification result of the training image comprises:
for each of the plurality of branches of the target re-identification network: processing the training image based on the convolutional layer of a branch to obtain a first feature information of the training image, the first feature information being configured as the feature information output by the branch.

6 . The method of claim 3 , wherein each of the plurality of branches comprises a convolutional layer, a sampling layer, and a classification layer; the identifying the training image based on the target re-identification network to obtain an identification result of the training image comprises:
for each of the plurality of branches of the target re-identification network: processing the training image based on the convolutional layer of a branch to obtain a first feature information of the training image; and processing the first feature information based on the sampling layer of the branch to obtain a second feature information of the training image, the second feature information being configured as the feature information output by the branch.

7 . The method of claim 3 , wherein each of the plurality of branches comprises a convolutional layer, a feature embedding layer, and a classification layer; the identifying the training image based on the target re-identification network to obtain an identification result of the training image comprises:
for each of the plurality of branches of the target re-identification network: processing the training image based on the convolutional layer of a branch to obtain a first feature information of the training image; and processing the first feature information based on the feature embedding layer of the branch to obtain a third feature information of the training image, the third feature information being configured as the feature information output by the branch.

8 . The method of claim 4 , wherein the number of the plurality of branches is n, each of the n branches comprises n convolutional layers; convolutional layers shared by different branches are different.

9 . The method of claim 4 , wherein each of the plurality of branches comprises further comprises a classification layer; the identifying the training image based on the target re-identification network to obtain an identification result of the training image further comprises:
processing the feature information output by the branch based on the classification layer of the branch to obtain a classification result of the branch; wherein the calculating a loss of each of the plurality of branches of the target re-identification network based on the identification result of the training image comprises: calculating the classification result of the branch based on a second loss function to obtain a second loss of the branch of the target re-identification network.

10 . The method according to claim 4 , wherein the first loss function is a triple loss function, and the second loss function is a cross-entropy loss function.

11 . The method of claim 8 , wherein:
the size of the feature information output by a last convolutional layer of a first branch of the n branches is smaller than the size of the feature information output by a last convolutional layer of other branches of the n branches; and/or, the n is 4, the n branches comprise a first branch, a second branch, a third branch, and a fourth branch; each of the 4 branches comprises a first convolutional layer, a second convolutional layer, a third convolutional layer, and a fourth convolutional layer;
the second branch and the first branch share the first convolutional layer and the second convolutional layer; the third branch and the second branch share the first convolutional layer, the second convolutional layer, and the third convolutional layer; the fourth branch and the first branch share the first convolutional layer, the second convolutional layer, and the third convolutional layer.

12 . The method of claim 9 , wherein the processing the feature information o

Assignees

• Zhejiang Dahua Technology Co

Inventors

• Sun Libo
• Pan Huadong
• Yin Jun

Classifications

• G06V2201/07

  Target detection · CPC title

• G06V10/82

  using neural networks · CPC title

• G06V20/30

  in albums, collections or shared content, e.g. social network photos or video · CPC title

• G06V10/454

  Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN] · CPC title

• G06V10/774 Primary

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