CN-122024274-A - Reloading pedestrian re-identification method based on semantic collaboration and chebyshev guide diagram reordering

CN122024274ACN 122024274 ACN122024274 ACN 122024274ACN-122024274-A

Abstract

The invention belongs to the technical field of reloading pedestrian re-identification, and particularly relates to a reloading pedestrian re-identification method based on semantic cooperation and chebyshev guide diagram reordering, which comprises a double-branch attention semantic cooperation module (ASMR), a frequency domain guide space division module (WGSD) and a diagram reordering module (Cheb-GR) based on chebyshev theorem, wherein an input image firstly extracts initial characteristics F through a Backbone network and sequentially inputs ASMR branches of two symmetrical structures to respectively model content semantics and obvious semantics related to identity. The method constructs a semantic synergetic module (ASMR) based on a double pooling strategy on the premise of keeping single RGB mode input, and is used for jointly mining content and remarkable semantic information, and introduces a space division module (WGSD) guided by a frequency domain, models a high-low frequency structure through wavelet decomposition, and realizes the selective enhancement of a discrimination area.

Inventors

ZHANG SHIJIA
ZHANG JIANXUN
ZENG YINGZHU

Assignees

重庆理工大学

Dates

Publication Date: 20260512
Application Date: 20251204

Claims (6)

1. The reloading pedestrian re-identification method based on semantic cooperation and chebyshev guide diagram reordering is characterized by comprising a double-branch attention semantic cooperation module (ASMR), a frequency domain guided space division module (WGSD) and a diagram reordering module (Cheb-GR) based on chebyshev theorem; firstly, extracting an initial characteristic F from an input image through a Backbone network, sequentially inputting ASMR branches of two symmetrical structures, and respectively modeling content semantics and salient semantics related to identity; Each ASMR branch comprises two sub-modules of semantic mining (SEMANTICMINING) and semantic optimization (SEMANTICREFINEMENT), wherein WGSD is embedded into the semantic mining module and used as a local blocking enhancement mechanism to realize frequency domain attention modeling based on wavelet decomposition; secondly, extracting semantic features from two semantic branches in each ASMR branch in the first step through average pooling and maximum pooling respectively, and performing cross fusion and fine granularity enhancement in a cross-branch architecture to finally obtain discrimination features with stronger semantic consistency The method is used for supervising the identity classification loss and the triplet loss in the training stage; Thirdly, wavelet transformation and frequency domain modeling are carried out, multi-scale frequency characteristics of a modeling image are obtained, computational efficient Haar wavelet transformation is adopted, the orthogonal base property of the Haar wavelet transformation can ensure lossless reconstruction of information, and in WGSD modules, characteristics output by a backstone are firstly obtained Performing Haar wavelet decomposition to obtain Wherein the low frequency component Preserving a global structure; Fourth, space blocking and local feature modeling is performed, WGSD, firstly, a frequency domain enhanced feature map is obtained through wavelet decomposition, the feature can keep a global structure and highlight fine-grained local information, on the basis, a module dynamically determines the boundary of a local area through an attention guiding mechanism, semantic alignment of a frequency domain layer is achieved, robustness of the feature is further enhanced under clothing change conditions, and in this way, WGSD can adaptively model multi-scale frequency information and improve capturing capability of the model on a discriminative local area. Acquiring frequency domain enhancement features Thereafter WGSD performs a horizontal division to extract region-specific characterizations, in particular, the feature map is divided into P groups of equal-width regions, each region Is performed by (1): ; Fifthly, based on ASMR modules of average pooling and maximum pooling, respectively denoted as ASMR-C and ASMR-S, which are respectively used for mining and refining identity-related semantic content and significant semantics; And finally, in order to improve the retrieval precision of the model in the test stage, a graph reordering method based on Chebyshev inequality is introduced after feature extraction to optimize the final retrieval result, and the method avoids the high calculation cost of the traditional k-nearest neighbor-based search by constructing a self-adaptive adjacency graph, and performs feature propagation in a graph structure by a parameter-free graph rolling mechanism, so that the context consistency of the retrieval result is effectively enhanced.
2. The method for reloading and identifying the pedestrian based on semantic cooperation and chebyshev guide diagram reordering according to claim 1, wherein in the third step, it is pointed out that when the LTCC data set is processed, the WGSD module performs two layers of cascade wavelet decomposition to obtain deeper frequency domain features, and only performs one layer of wavelet decomposition on the PRCC and Celeb-reID data sets to achieve both efficiency and performance; ; ; Wherein the method comprises the steps of As a high-frequency enhancement factor that can be learned, Representing the low frequency characteristics of bilinear upsampling recovery.
3. The reloading pedestrian re-identification method based on semantic collaboration and chebyshev guide diagram reordering according to claim 1, wherein in the fourth step Respectively represent the first The height and width of each region and the frequency domain enhancement characteristic of the region are projected through the shared full-connection layer, and the obtained local characteristic is fused with the global characteristic and finally identified.
4. The reloading pedestrian re-identification method based on semantic cooperation and Chebyshev guide diagram reordering according to claim 1, wherein ASMR-C and ASMR-S are alternately arranged in two branches, the sequences are different but complement each other, on one hand, the content semantic features after refining The method is favorable for further mining the significant semantic information and finally obtaining the dual-refined significant semantic features The loss function of the first branch is calculated as follows: ; on the other hand, refined salient semantic features Further mining of content semantic information is facilitated, and finally double-refined content semantic features are obtained The loss function of the second branch is calculated as follows: 。
5. The reloading pedestrian re-recognition method based on semantic collaboration and chebyshev guide diagram reordering of claim 4, wherein features of two branches are respectively focused on content and salient semantics, and are fused into one comprehensive semantic feature, expressed as follows: 。
6. the reloading pedestrian re-identification method based on semantic collaboration and chebyshev guide diagram reordering as claimed in claim 5, wherein the method is characterized in that Representing element-by-element addition, conv represents a convolution block, the scheme pooling global maximization As final pedestrian re-identification identity representation and supervised by a usual pedestrian re-identification loss function, in particular: ; ; wherein MaxPool denotes a global max pooling operation.

Description

Reloading pedestrian re-identification method based on semantic collaboration and chebyshev guide diagram reordering Technical Field The invention belongs to the technical field of reloading pedestrian re-identification, and particularly relates to a reloading pedestrian re-identification method based on semantic cooperation and chebyshev guide diagram reordering. Background Pedestrian re-identification (PersonRe-identification, personRe-ID) is intended to identify a particular pedestrian across cameras in non-overlapping fields of view. The method has important application value in the fields of smart cities, monitoring safety and the like, and has received a great deal of attention in recent years. Although the traditional Re-ID method achieves remarkable effect [1-5] in a short-term scene with unchanged clothes, the core assumption that target characters are kept on the same dressing under different camera angles severely restricts the applicability of the method in an actual long-term monitoring scene. To break through this limitation, researchers have recently presented a more challenging task of reloading pedestrian re-identification (CCRe-ID), aimed at developing identification models that are robust to changes in clothing to meet the urgent need for long-term monitoring in the real world. Existing reloader re-identification studies have proposed various strategies to address challenges due to significant changes in appearance. One class of methods adopts a generated model to synthesize images of the same pedestrian under different clothes to learn the identity characteristics irrelevant to the clothes, and the other class of methods focuses on introducing multi-modal information such as key points, outlines, gait or three-dimensional body types of auxiliary characteristics of human bodies so as to enhance the robustness of the identity characterization. However, the generating method has high requirements on image quality and high training cost, the multi-mode method relies on a plurality of auxiliary models to work cooperatively, the calculation complexity is high, the performance is easily affected by the precision of each mode, the actual deployment difficulty is high, and the application of the method in a real scene is limited. Disclosure of Invention The invention aims to provide a reloading pedestrian re-identification method based on semantic collaboration and chebyshev guide diagram reordering, and aims to solve the problems that the current reloading pedestrian re-identification generation method has high requirements on image quality and high training cost, the multi-mode method relies on a plurality of auxiliary models to work cooperatively, the calculation complexity is high, the performance is easily influenced by the precision of each mode, the actual deployment difficulty is high, and the application of the method in a real scene is limited. In order to achieve the purpose, the invention provides the technical scheme that the reloading pedestrian re-identification method based on semantic cooperation and chebyshev guide diagram reordering comprises a double-branch attention semantic cooperation module (ASMR), a frequency domain guide space division module (WGSD) and a chebyshev theorem-based diagram reordering module (Cheb-GR); firstly, extracting an initial characteristic F from an input image through a Backbone network, sequentially inputting ASMR branches of two symmetrical structures, and respectively modeling content semantics and salient semantics related to identity; Each ASMR branch comprises two sub-modules of semantic mining (SEMANTICMINING) and semantic optimization (SEMANTICREFINEMENT), wherein WGSD is embedded into the semantic mining module and used as a local blocking enhancement mechanism to realize frequency domain attention modeling based on wavelet decomposition; secondly, extracting semantic features from two semantic branches in each ASMR branch in the first step through average pooling and maximum pooling respectively, and performing cross fusion and fine granularity enhancement in a cross-branch architecture to finally obtain discrimination features with stronger semantic consistency The method is used for supervising the identity classification loss and the triplet loss in the training stage; Thirdly, wavelet transformation and frequency domain modeling are carried out, multi-scale frequency characteristics of a modeling image are obtained, computational efficient Haar wavelet transformation is adopted, the orthogonal base property of the Haar wavelet transformation can ensure lossless reconstruction of information, and in WGSD modules, characteristics output by a backstone are firstly obtained Performing Haar wavelet decomposition to obtainWherein the low frequency componentPreserving a global structure; Fourthly, modeling the space block and the local feature, WGSD firstly obtaining a frequency domain enhancement feature map through wavelet decomposition, wh