Title:
PERSON IMAGE RE-IDENTIFICATION METHOD BASED ON AUTONOMOUS MODEL STRUCTURE EVOLUTION
Document Type and Number:
WIPO Patent Application WO/2024/093466
Kind Code:
A1
Abstract:
Disclosed in the present invention is a person image re-identification method based on autonomous model structure evolution. A main structure is divided into three functional modules: a basic feature extraction module, a self-evolution multi-scale feature enhancement module and a multi-part fine-grained alignment module. In the basic feature extraction module, a basic feature of an image is extracted by using a visual convolutional neural network; thereafter, the self-evolution multi-scale feature enhancement module is inserted into the basic feature extraction module, a dynamic routing mechanism is used for being responsible for multi-scale visual feature enhancement, and by using deep and shallow features extracted by the basic feature extraction module, and by means of layer-by-layer scale transform and feature refining, a model is guided to learn how to evolve from a basic visual feature to an adaptive multi-scale feature; and finally, in the multi-part fine-grained alignment module, the model is guided to extract person features of multiple parts by using external semantic knowledge, and the matching of local features is realized. By means of the present invention, the accuracy of a local person image retrieval task can be improved.
Inventors:
ZHANG YANNING (CN)
WANG PENG (CN)
CHEN HONGYU (CN)
JIAO BINGLIANG (CN)
GAO LIYING (CN)
WANG PENG (CN)
CHEN HONGYU (CN)
JIAO BINGLIANG (CN)
GAO LIYING (CN)
Application Number:
PCT/CN2023/114843
Publication Date:
May 10, 2024
Filing Date:
August 25, 2023
Export Citation:
Assignee:
NORTHWESTERN POLYTECHNICAL UNIV (CN)
International Classes:
G06V40/10
Attorney, Agent or Firm:
ZHENGZHOU YUYUAN INTELLECTUAL PROPERTY AGENCY OFFICE (GENERAL PARTNERSHIP) (12th Floor Building 1, No. 49, Jinshui East Road, Zhengzhou Area , Henan Pilot Free Trade Zon, Zhengzhou Henan 0, CN)
Download PDF: