A cross-view protective clothing feature matching method and system based on contrast learning

Abstract

The application relates to the technical field of image recognition, and discloses a cross-view protective clothing feature matching method and system based on contrast learning. The method comprises the following steps: cross-view acquisition and processing of protective clothing to obtain standard cross-view protective clothing images; embedding a view self-adaptive coding into a contrast learning feature extraction network to correct distortion and extract features of the standard cross-view protective clothing acquisition images, and obtaining self-adaptive initial features of the protective clothing; constructing a positive and negative sample pair of contrast learning, optimizing spatial mapping of the contrast learning feature extraction network, and obtaining cross-view alignment features of the protective clothing; performing feature enhancement processing on the cross-view alignment features based on a twin coding encoder combined with an attention mechanism, and obtaining enhanced matching features of the protective clothing; and performing matching confirmation on the enhanced matching features, and outputting a cross-view feature matching result of the protective clothing, so that the representation stability and matching accuracy of the cross-view features of the protective clothing are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a cross-view feature matching method and system for protective clothing based on contrastive learning. Background Technology

[0002] As critical protective equipment in special scenarios, cross-view feature matching is a core technical step in achieving equipment identification, verification, and management of protective suits. In practical applications, multi-view image acquisition of protective suits is necessary, relying on feature extraction and matching technologies to complete target determination. The contrastive learning paradigm provides a fundamental technical path for image feature processing. The combined application of view-adaptive coding, twin encoders, and attention mechanisms also provides technical directions for optimizing the cross-view feature processing flow. Developing cross-view feature matching methods adapted to protective suits has become an important requirement for improving the accuracy of equipment management.

[0003] Existing cross-view protective clothing feature matching technologies suffer from several shortcomings. They fail to effectively integrate viewpoint adaptive coding with the feature extraction process, resulting in an inability to accurately correct distortion in cross-view acquired images and leading to a lack of stability in the extracted features. Furthermore, they fail to construct dedicated sample pairs based on the correlation between cross-view protective clothing images, resulting in inadequate feature space mapping optimization. The twin encoder and attention mechanism are not properly applied, leading to poor feature enhancement. Finally, the matching confirmation stage lacks rigorous verification logic, resulting in low accuracy and poor reliability of the final cross-view feature matching results, which cannot meet the practical application requirements for precise matching of protective clothing. Summary of the Invention

[0004] This invention provides a cross-view protective clothing feature matching method and system based on contrastive learning to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, this invention provides a cross-view protective clothing feature matching method based on contrastive learning, comprising: C1: Perform cross-view acquisition processing on the protective suit to obtain a standard cross-view protective suit image, and input the standard cross-view protective suit acquired image into the contrast learning feature extraction network; C2: Embed the viewpoint adaptive coding into the contrastive learning feature extraction network to perform distortion correction and feature extraction on the standard cross-view protective clothing images to obtain the adaptive initial features of the protective clothing; C3: Based on the correlation between the cross-view protective clothing images, construct contrastive learning positive and negative sample pairs, and based on the contrastive learning positive and negative sample pairs, perform spatial mapping optimization on the contrastive learning feature extraction network to obtain the cross-view alignment features of the protective clothing; C4: Based on the twin encoder combined with the attention mechanism, the cross-view alignment features are enhanced to obtain the enhanced matching features of the protective suit; C5: Confirm the matching of the enhanced matching features and output the cross-view feature matching result of the protective suit.

[0006] In a preferred embodiment, the step of performing cross-view acquisition processing on the protective suit to obtain a standard cross-view protective suit image, and inputting the standard cross-view protective suit acquired image into a contrastive learning feature extraction network, includes: The original images of the protective suit acquired from multiple perspectives are screened for viewpoint consistency, and image data that does not match the target acquisition perspective are removed to obtain the protective suit image set. Image quality normalization is performed on the set of protective clothing images to obtain standard cross-view protective clothing images of the protective clothing; Based on preset features, the standard cross-view protective clothing image is format adapted, and the format adaptation result is transmitted to the feature input port of the contrastive learning feature extraction network.

[0007] In a preferred embodiment, embedding the viewpoint adaptive coding into a contrastive learning feature extraction network to perform distortion correction and feature extraction on the standard cross-view protective clothing images to obtain the adaptive initial features of the protective clothing includes: Based on the perspective distortion attribute features in the standard cross-view protective clothing image, a perspective distortion attribute feature set is constructed, and the perspective distortion attribute feature set is customized and encoded to obtain the perspective adaptive encoding of the protective clothing. The viewpoint adaptive coding is connected to the feature preprocessing node of the contrastive learning feature extraction network to complete the adaptation and fusion of the viewpoint adaptive coding and the contrastive learning feature extraction network. Based on the aforementioned viewpoint adaptive coding, viewpoint distortion recognition is performed on the standard cross-viewpoint protective clothing image to obtain the distortion feature localization result of the protective clothing; Based on the distortion feature localization results, distortion correction and feature extraction are performed on the standard cross-view protective clothing image to obtain the adaptive initial features of the protective clothing.

[0008] In a preferred embodiment, the step of integrating the viewpoint adaptive coding into the feature preprocessing node of the contrastive learning feature extraction network to complete the adaptation and fusion of the viewpoint adaptive coding and the contrastive learning feature extraction network includes: Locate the feature preprocessing node of the contrastive learning feature extraction network to determine the access position of the viewpoint adaptive coding; Based on the access location, a data path connection is established between the viewpoint adaptive coding and the feature preprocessing node to create a transmission link between the viewpoint adaptive coding and the feature preprocessing node. Based on the transmission link, the viewpoint adaptive coding and the contrastive learning feature extraction network are adapted in a coordinated manner to achieve the adaptation and fusion of the viewpoint adaptive coding and the contrastive learning feature extraction network.

[0009] In a preferred embodiment, the step of constructing contrastive learning positive and negative sample pairs based on the correlation relationships of the cross-view protective clothing images, and then optimizing the spatial mapping of the contrastive learning feature extraction network based on the contrastive learning positive and negative sample pairs to obtain the cross-view alignment features of the protective clothing, includes: Based on the viewpoint association attributes of the cross-view protective clothing images, the same-target cross-view samples and different-target samples in the cross-view protective clothing images are distinguished, and the contrastive learning positive and negative sample pairs of the cross-view protective clothing images are constructed. Based on the contrastive learning positive and negative sample pairs, feature mapping orientation constraints are applied to the contrastive learning feature extraction network to obtain the optimized feature mapping process for the protective suit. Based on the optimized feature mapping process, directional feature mapping processing is performed on the adaptive initial features of the protective suit to generate cross-view alignment features of the protective suit.

[0010] In a preferred embodiment, the process of applying feature mapping orientation constraints to the contrastive learning feature extraction network based on the contrastive learning positive and negative sample pairs to obtain the optimized feature mapping of the protective suit includes: Based on the set of feature association parameters of the contrastive learning positive and negative sample pairs, the feature mapping constraint loss of the contrastive learning feature extraction network is calculated. Based on the feature mapping constraint loss, the feature mapping weights of the contrastive learning feature extraction network are re-adjusted, and the feature mapping path of the contrastive learning feature extraction network is normalized based on the re-adjustment, resulting in the optimized feature mapping process of the protective suit.

[0011] In a preferred embodiment, the formula for calculating the feature mapping constraint loss is: ; In the formula, This represents the feature mapping constraint loss of the protective suit. Indicates the temperature coefficient. This indicates the initial characteristics of the target protective suit. This represents the cross-view sample features of the protective suit targeting the same object. This indicates the heterogeneous target sample characteristics of the protective suit. This represents the cross-view sample association weight of the protective suit. The cosine similarity function is used. Let be the natural logarithm function with base e. It is an exponential function with the natural constant e as its base.

[0012] In a preferred embodiment, the step of performing feature enhancement processing on the cross-view alignment features based on a twin encoder combined with an attention mechanism to obtain the enhanced matching features of the protective suit includes: The cross-view alignment features are input into the twin encoder, and the cross-view alignment features are synchronously parsed through the dual-branch structure of the twin encoder to obtain the dual-branch synchronous parsing features of the protective suit. Extract the representation dimension and distribution information of the cross-view alignment features, and construct an attention mechanism that adapts to the dual-branch synchronous parsing features based on the representation rules of the core features of the protective suit. Based on the attention mechanism, key feature filtering is performed on the dual-branch synchronous parsing features to locate the core representation information of the protective clothing in the dual-branch synchronous parsing features. Based on the core representation information, the dual-branch synchronous parsing features are weighted and integrated to obtain the enhanced matching features of the protective suit.

[0013] In a preferred embodiment, the step of confirming the matching of the enhanced matching features and outputting the cross-view feature matching result of the protective suit includes: The enhanced matching features are subjected to bidirectional feature matching verification, and the feature matching confidence is determined based on the bidirectional feature matching verification results. Based on the feature matching confidence, the enhanced matching features are effectively filtered to obtain the target matching features of the protective suit. Based on the target matching features, the cross-view feature matching results of the protective suit are generated and output.

[0014] To address the aforementioned problems, this invention also provides a cross-view protective clothing feature matching system based on contrastive learning. The system includes a cross-view acquisition module, an adaptive feature extraction module, a spatial mapping optimization module, a feature enhancement processing module, and a feature matching confirmation module, wherein: The cross-view acquisition module is used to perform cross-view acquisition processing on the protective suit to obtain a standard cross-view protective suit image, and input the standard cross-view protective suit acquisition image into the contrast learning feature extraction network. The adaptive feature extraction module is used to embed the viewpoint adaptive coding into the contrastive learning feature extraction network, perform distortion correction and feature extraction on the standard cross-view protective clothing images, and obtain the adaptive initial features of the protective clothing. The spatial mapping optimization module is used to construct contrastive learning positive and negative sample pairs based on the correlation relationship of the cross-view protective clothing images, and to perform spatial mapping optimization on the contrastive learning feature extraction network based on the contrastive learning positive and negative sample pairs to obtain the cross-view alignment features of the protective clothing. The feature enhancement processing module is used to perform feature enhancement processing on the cross-view alignment features based on the twin encoder combined with the attention mechanism to obtain the enhanced matching features of the protective suit. The feature matching confirmation module is used to confirm the matching of the enhanced matching features and output the cross-view feature matching results of the protective suit.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention, by embedding viewpoint adaptive coding into a contrastive learning feature extraction network, enables precise distortion correction and feature extraction of standard cross-view protective clothing images. It effectively eliminates image distortion interference caused by cross-view acquisition, generating adaptive initial features that fit the characteristics of the protective clothing itself, ensuring the stability and integrity of the initial features. By constructing contrastive learning positive and negative sample pairs based on the correlation between cross-view protective clothing images, it achieves directional spatial mapping optimization of the contrastive learning feature extraction network, allowing for unified alignment of protective clothing features from different viewpoints. This results in cross-view aligned features with stronger representational capabilities, improving the universality and recognizability of features from the source, providing high-quality basic data for subsequent feature processing, and ensuring the coherence and rationality of the feature processing flow.

[0016] 2. This invention employs a twin encoder combined with an attention mechanism to enhance cross-view alignment features. This allows for precise focus on the core representational information of protective clothing, strengthening the expression of key features and generating more recognizable enhanced matching features, further improving the matching value of the features. Through rigorous matching verification of the enhanced matching features, accurate and reliable cross-view feature matching results can be output, significantly improving the accuracy and reliability of cross-view feature matching for protective clothing. The technical solution of this invention features a tightly integrated process, with each step working synergistically. It can adapt to the cross-view feature matching needs of protective clothing in various scenarios, effectively improving the efficiency of protective clothing identification and verification, providing stable technical support for the standardized management of protective clothing, and possessing strong practical value and application prospects. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a cross-view protective clothing feature matching method based on contrastive learning, provided in an embodiment of the present invention. Figure 2 A functional block diagram of a cross-view protective clothing feature matching system based on contrastive learning, provided in an embodiment of the present invention; The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] This application provides a cross-view feature matching method for protective clothing based on contrastive learning. The executing entity of this cross-view feature matching method for protective clothing based on contrastive learning includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, this cross-view feature matching method for protective clothing based on contrastive learning can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0020] Reference Figure 1 The diagram shown is a flowchart illustrating a cross-view protective clothing feature matching method based on contrastive learning according to an embodiment of the present invention. In this embodiment, the cross-view protective clothing feature matching method based on contrastive learning includes: C1: Perform cross-view acquisition processing on the protective suit to obtain a standard cross-view protective suit image, and input the standard cross-view protective suit acquired image into the contrast learning feature extraction network; In this embodiment of the invention, the step of performing cross-view acquisition processing on the protective suit to obtain a standard cross-view protective suit image, and inputting the standard cross-view protective suit acquired image into a contrastive learning feature extraction network, includes: The original images of the protective suit acquired from multiple perspectives are screened for viewpoint consistency, and image data that does not match the target acquisition perspective are removed to obtain the protective suit image set. Image quality normalization is performed on the set of protective clothing images to obtain standard cross-view protective clothing images of the protective clothing; Based on preset features, the standard cross-view protective clothing image is format adapted, and the format adaptation result is transmitted to the feature input port of the contrastive learning feature extraction network.

[0021] It should be specifically explained that, according to the preset target acquisition angle, all original images of the protective suit acquired from multiple perspectives are subjected to perspective matching and verification one by one. Image data whose perspective does not match the target acquisition angle are all removed, and image data that meets the perspective requirements are retained and summarized to form the protective suit image set.

[0022] Furthermore, the preset target acquisition perspective refers to the combination of standardized protective clothing observation angles, shooting positions, and acquisition postures that are pre-set to ensure that the protective clothing images acquired from multiple perspectives have a unified perspective benchmark. This includes pre-determined core observation angles such as the front, side, and back of the protective clothing, as well as fixed shooting distances, shooting heights, and shooting position conditions at the corresponding angles. These are used as a unified judgment basis for screening original images from multiple perspectives, ensuring that the screened images all meet the unified perspective acquisition standards.

[0023] All images within the protective suit image set are sequentially processed for sharpness optimization, brightness equalization adjustment, and size specification unification to eliminate image quality differences and specification deviations between images within the set, thereby generating a standard cross-view protective suit image.

[0024] Furthermore, the standard cross-view protective clothing image is a standardized protective clothing imaging data formed after the viewpoint consistency screening, image quality regularization and format adaptation. It provides a unified benchmark image carrier for subsequent distortion correction, feature extraction and matching processing. It includes complete imaging images of the protective clothing from pre-set core observation views such as the front, side and back. All viewpoint images are acquired using a uniform shooting distance, shooting height and shooting orientation, without viewpoint shift, imaging defects or local occlusion. It completely preserves all appearance characterization information of the protective clothing, such as the overall outline, fabric texture, identification patterns, structural splicing positions, and protective component shapes, and can fully support the core feature acquisition required for subsequent feature extraction and matching.

[0025] Based on the input receiving requirements of the contrastive learning feature extraction network, the standard cross-view protective clothing image is format converted and feature dimension adapted, and the adapted standard cross-view protective clothing image is transmitted to the feature input port of the contrastive learning feature extraction network.

[0026] Furthermore, the contrastive learning feature extraction network is an integrated feature processing architecture built for cross-view protective clothing image feature processing. It consists of a feature input port, a feature preprocessing node, a feature mapping processing unit, and a feature output port connected in sequence. Each component works together to complete the entire process of receiving, preprocessing, extracting features, and transmitting protective images.

[0027] Among them, the feature input port is used to receive standard cross-view protective clothing images adapted in format and dimension, establish a stable transmission path for image data, and avoid image data loss or distortion during the transmission process.

[0028] The feature preprocessing node provides a dedicated embedding location for viewpoint adaptive coding, which can complete the adaptation and fusion of coding and network. Based on coding, it realizes viewpoint distortion recognition and correction of standard cross-view protective clothing images, and eliminates image distortion interference caused by cross-view acquisition.

[0029] The feature mapping processing unit is used to perform directional feature extraction on the distortion-corrected protective clothing image. It can adjust the feature mapping processing logic and path according to external constraints to make the extracted features fit the protective clothing's properties and adapt to the feature distribution patterns under different viewpoints.

[0030] The feature output port is used to transmit the extracted adaptive initial features outward, providing standardized basic feature data for subsequent feature space mapping optimization and feature enhancement processing, ensuring that the feature source for subsequent processing steps is stable and consistent.

[0031] By performing viewpoint consistency screening and image quality regularization on the original images of protective suits acquired from multiple perspectives, standard cross-view protective suit images with unified viewpoints, balanced image quality, and standardized specifications can be formed. This effectively avoids problems such as viewpoint shift and inconsistent imaging quality caused by multi-view acquisition. At the same time, relying on format adaptation processing, it ensures that the images can be stably connected to the contrastive learning feature extraction network, providing standardized and highly recognizable basic image data for subsequent distortion correction, feature extraction, and spatial mapping optimization, ensuring the continuity and stability of the subsequent feature processing flow.

[0032] C2: Embed the viewpoint adaptive coding into the contrastive learning feature extraction network to perform distortion correction and feature extraction on the standard cross-view protective clothing images to obtain the adaptive initial features of the protective clothing; In this embodiment of the invention, the step of embedding the viewpoint adaptive coding into a contrastive learning feature extraction network to perform distortion correction and feature extraction on the standard cross-view protective clothing images to obtain the adaptive initial features of the protective clothing includes: Based on the perspective distortion attribute features in the standard cross-view protective clothing image, a perspective distortion attribute feature set is constructed, and the perspective distortion attribute feature set is customized and encoded to obtain the perspective adaptive encoding of the protective clothing. The viewpoint adaptive coding is connected to the feature preprocessing node of the contrastive learning feature extraction network to complete the adaptation and fusion of the viewpoint adaptive coding and the contrastive learning feature extraction network. Based on the aforementioned viewpoint adaptive coding, viewpoint distortion recognition is performed on the standard cross-viewpoint protective clothing image to obtain the distortion feature localization result of the protective clothing; Based on the distortion feature localization results, distortion correction and feature extraction are performed on the standard cross-view protective clothing image to obtain the adaptive initial features of the protective clothing.

[0033] The step of integrating the viewpoint adaptive coding into the feature preprocessing node of the contrastive learning feature extraction network to complete the adaptation and fusion of the viewpoint adaptive coding and the contrastive learning feature extraction network includes: Locate the feature preprocessing node of the contrastive learning feature extraction network to determine the access position of the viewpoint adaptive coding; Based on the access location, a data path connection is established between the viewpoint adaptive coding and the feature preprocessing node to create a transmission link between the viewpoint adaptive coding and the feature preprocessing node. Based on the transmission link, the viewpoint adaptive coding and the contrastive learning feature extraction network are adapted in a coordinated manner to achieve the adaptation and fusion of the viewpoint adaptive coding and the contrastive learning feature extraction network.

[0034] It should be specifically explained that all view distortion attribute features are extracted from standard cross-view protective clothing images and summarized to form a view distortion attribute feature set. According to the distortion characterization rules of cross-view imaging of protective clothing, the view distortion attribute feature set is subjected to customized coding processing to generate view adaptive coding of protective clothing.

[0035] Furthermore, the set of viewpoint distortion attribute features includes imaging orientation shift features, geometric distortion features, shooting viewpoint deviation features, image edge stretching distortion features, overall outline proportion distortion features of the protective suit, imaging deformation features of local protective components, as well as imaging angle deviation, size scaling distortion, and regional position shift under different observation angles. All attribute features can accurately reflect the degree, distribution location and type of distortion. All types of features are interrelated and fully present the whole picture of image viewpoint distortion, ensuring that the generated viewpoint adaptive code can fully adapt to the distortion representation requirements of cross-viewpoint imaging of protective suits.

[0036] The location of feature preprocessing nodes is located and determined within the internal architecture of the contrastive learning feature extraction network, thereby clarifying the access point for viewpoint adaptive encoding and ensuring that the encoding access point matches the image preprocessing process.

[0037] Furthermore, the viewpoint adaptive coding is a dedicated processing code generated by custom coding integration and feature adaptation adjustment based on the viewpoint distortion attribute feature set of standard cross-viewpoint protective clothing images. It is specifically designed to meet the needs of distortion correction and feature extraction of cross-viewpoint protective clothing images and is the core carrier for realizing viewpoint adaptive processing.

[0038] The viewpoint adaptive coding specifically includes distortion feature matching sub-coding, network parameter adaptation sub-coding, distortion correction execution sub-coding, and feature extraction guidance sub-coding. Among them, the distortion feature matching sub-coding collects all viewpoint distortion attribute features, which can quickly identify the type, distribution location, and degree of distortion in the image. The network parameter adaptation sub-coding carries the interface parameters and operating specifications that match the feature preprocessing nodes of the contrastive learning feature extraction network, ensuring stable connection between the coding and the network and smooth data transmission. The distortion correction execution sub-coding has built-in execution logic for geometric correction, contour restoration, and scale calibration of the protective clothing image, which can complete distortion elimination processing in a targeted manner based on the distortion localization results. The feature extraction guidance sub-coding marks the extraction direction of the core appearance features of the protective clothing, which can accurately extract effective features and remove distortion interference features after distortion correction. All sub-codings work together to fully realize the adaptive processing of the entire process from distortion recognition and correction to feature extraction.

[0039] Based on the determined access location, the output end of the viewpoint adaptive coding is connected to the input end of the feature preprocessing node to establish a stable data transmission link between the viewpoint adaptive coding and the feature preprocessing node.

[0040] Based on the established transmission link, the operating parameters of the viewpoint adaptive coding and the preprocessing parameters of the contrastive learning feature extraction network are synchronously matched and adjusted to ensure that the parameter specifications of the coding and the network are consistent, thus completing the adaptation and fusion of viewpoint adaptive coding and contrastive learning feature extraction network.

[0041] Furthermore, the preprocessing parameters of the contrastive learning feature extraction network specifically include code interface adaptation parameters, image distortion detection parameters, image correction execution parameters, basic feature extraction parameters, and data synchronization and coordination parameters. Among them, the code interface adaptation parameters are used to standardize the access transmission specifications, data format, and signal matching standards of the viewpoint adaptive coding, ensuring stable connection between the coding and preprocessing nodes. The image distortion detection parameters are used to set the distortion scanning range, distortion judgment rules, and feature positioning accuracy of the protective clothing image, working in conjunction with the viewpoint adaptive coding to achieve accurate distortion recognition. The image correction execution parameters are used to determine the execution standards for image geometric correction, contour restoration, and scaling, providing a unified basis for distortion correction. The basic feature extraction parameters are used to limit the feature extraction area, feature selection rules, and output specifications, standardizing the extraction standards for adaptive initial features. The data synchronization and coordination parameters are used to coordinate the processing timing and running rhythm of the coding and network, ensuring that the coding output and network preprocessing operations are executed synchronously. The overall parameter system fully supports the preprocessing nodes in completing the entire process of coding fusion, distortion recognition, correction, and feature extraction.

[0042] The steps of synchronous matching and adjustment are as follows: extract the running parameters of the viewpoint adaptive coding to form a coding adaptation parameter set; retrieve the preprocessing parameters of the network feature preprocessing node of the contrast learning feature extraction network to form a network baseline parameter set; compare and verify the two parameter sets item by item and locate the mismatched parameter items; use the network baseline parameter set as the standard to perform targeted correction on the mismatched items of the coding adaptation parameter set; and then carry out linkage operation verification on the corrected parameters.

[0043] By using fused viewpoint adaptive coding, the imaging area of ​​the standard cross-view protective clothing image is scanned and detected in its entirety. The viewpoint distortion areas in the image are identified and the distortion locations are marked, thus obtaining the distortion feature localization results of the protective clothing.

[0044] Furthermore, the distortion feature localization result is precise calibration information generated after distortion identification of standard cross-view protective clothing images by viewpoint adaptive coding, providing a direct basis for subsequent distortion correction. This result specifically includes the coordinate position, distortion type, distortion degree, and distribution range of distorted regions in the image, as well as the distortion offset and deformation amplitude information of key parts such as the protective clothing outline, protective components, and structural joints. It fully marks the specific location and impact of distortion problems such as imaging orientation shift, geometric distortion, proportional distortion, and edge stretching.

[0045] Distortion correction processing is performed on the standard cross-view protective clothing image based on the distortion region marked by the distortion feature localization result to eliminate the view distortion problem in the image. Then, the exclusive appearance features of the protective clothing are extracted from the corrected image to obtain the adaptive initial features of the protective clothing.

[0046] Furthermore, the adaptive initial features fully include the overall outline features, fabric texture features, identification pattern features, structural splicing features, protective component morphological features, and size ratio features of the protective suit. At the same time, the standard coordinate position, distribution pattern, and associated attributes of each feature in the image are marked. All features have been freed from view distortion interference, maintaining a unified view reference and representation standard, which can truly reflect the appearance attributes and structural features of the protective suit itself.

[0047] By customizing the viewpoint distortion attribute features based on standard cross-view protective clothing images to generate viewpoint adaptive codes, and then adapting and fusing the codes with a contrastive learning feature extraction network, the distorted regions in the images can be accurately identified and located based on the codes, effectively eliminating the imaging distortion interference caused by multi-view acquisition. Then, the core representation features of the protective clothing, after removing the distortion effects, are extracted from the corrected images to obtain standardized and unified adaptive initial features. This provides real and reliable basic feature data for subsequent feature mapping optimization and feature enhancement processing, greatly improving the accuracy and adaptability of cross-view protective clothing feature extraction.

[0048] C3: Based on the correlation between the cross-view protective clothing images, construct contrastive learning positive and negative sample pairs, and based on the contrastive learning positive and negative sample pairs, perform spatial mapping optimization on the contrastive learning feature extraction network to obtain the cross-view alignment features of the protective clothing; In this embodiment of the invention, the step of constructing contrastive learning positive and negative sample pairs based on the correlation relationship of the cross-view protective clothing images, and optimizing the spatial mapping of the contrastive learning feature extraction network based on the contrastive learning positive and negative sample pairs to obtain the cross-view alignment features of the protective clothing, includes: Based on the viewpoint association attributes of the cross-view protective clothing images, the same-target cross-view samples and different-target samples in the cross-view protective clothing images are distinguished, and the contrastive learning positive and negative sample pairs of the cross-view protective clothing images are constructed. Based on the contrastive learning positive and negative sample pairs, feature mapping orientation constraints are applied to the contrastive learning feature extraction network to obtain the optimized feature mapping process for the protective suit. Based on the optimized feature mapping process, directional feature mapping processing is performed on the adaptive initial features of the protective suit to generate cross-view alignment features of the protective suit.

[0049] The process of applying feature mapping orientation constraints to the contrastive learning feature extraction network based on the contrastive learning positive and negative sample pairs to obtain the optimized feature mapping process for the protective suit includes: Based on the set of feature association parameters of the contrastive learning positive and negative sample pairs, the feature mapping constraint loss of the contrastive learning feature extraction network is calculated. Based on the feature mapping constraint loss, the feature mapping weights of the contrastive learning feature extraction network are re-adjusted, and the feature mapping path of the contrastive learning feature extraction network is normalized based on the re-adjustment, resulting in the optimized feature mapping process of the protective suit.

[0050] The formula for calculating the feature mapping constraint loss is as follows: ; In the formula, This represents the feature mapping constraint loss of the protective suit. Indicates the temperature coefficient. This indicates the initial characteristics of the target protective suit. This represents the cross-view sample features of the protective suit targeting the same object. This indicates the heterogeneous target sample characteristics of the protective suit. This represents the cross-view sample association weight of the protective suit. The cosine similarity function is used. Let be the natural logarithm function with base e. It is an exponential function with the natural constant e as its base.

[0051] It needs to be specifically explained that all images are classified and divided based on the perspective association attributes between cross-view protective clothing images. Images belonging to the same protective clothing but different observation perspectives are defined as cross-view samples of the same target, and images belonging to different protective clothing are defined as samples of different targets. According to the rule that positive samples correspond to cross-view samples of the same target and negative samples correspond to samples of different targets, a comparative learning positive and negative sample pair of cross-view protective clothing images is constructed.

[0052] Furthermore, the contrastive learning positive and negative sample pairs are feature constraint sample combinations constructed to achieve cross-view protective clothing feature space mapping optimization. They consist of positive and negative sample pairs and serve as the core basis for implementing feature mapping orientation constraints on the contrastive learning feature extraction network.

[0053] Positive samples are cross-viewpoint samples of the same target, specifically referring to standard cross-viewpoint images of protective suits obtained from the same protective suit under different observation positions at a preset standard acquisition angle. These samples represent the appearance features of the same protective suit presented from different perspectives and have the attributes of feature homology and high correlation. Negative samples are heterogeneous samples, specifically referring to standard cross-viewpoint images of protective suits from different protective suits, regardless of whether the perspective is consistent. These samples represent the differentiated appearance features of different protective suits and have no feature homology with the target samples.

[0054] This sample provides clear guidance for network feature mapping through the homology association constraint of positive samples and the heterology distinction constraint of negative samples, enabling the network to learn the unified spatial distribution pattern of cross-view features of the same protective suit, laying the foundation for subsequent feature alignment and optimization.

[0055] The feature association parameters between each sample within the positive and negative sample pairs of the contrastive learning are extracted and summarized to form a feature association parameter set. Combining the feature matching relationship and the degree of association between the samples, the feature mapping constraint loss of the contrastive learning feature extraction network is determined. This loss is used to characterize the degree of deviation between the network feature mapping result and the constraint requirements of the positive and negative samples.

[0056] Furthermore, the feature association parameter set is a collection of core parameters extracted from positive and negative sample pairs in comparative learning. These parameters are used to quantify the feature association attributes and matching degree between samples, and serve as the key data basis for calculating feature mapping constraint loss and implementing feature mapping directional constraints. This set includes feature homology association parameters, feature similarity matching parameters, cross-view feature correspondence parameters, and cross-view sample association weight parameters between samples with the same target across different perspectives; feature heterogeneity discrimination parameters, feature difference determination parameters, and sample irrelevance calibration parameters between samples with different targets; and feature association strength parameters, feature distribution deviation parameters, and feature mapping matching benchmark parameters between target samples and positive samples, and target samples and negative samples. All these parameters comprehensively reflect the tightness, difference, and correspondence of feature associations within positive and negative sample pairs, providing accurate data support for calculating feature mapping constraint loss and adjusting network weights.

[0057] Feature mapping constraint loss is a core metric for network optimization. The larger the value, the greater the deviation between the network feature mapping result and the constraint requirements. The smaller the value, the better the mapping effect fits the cross-view feature alignment requirements. It is used to directly guide the adjustment direction of feature mapping weights and paths.

[0058] The temperature coefficient is a preset fixed hyperparameter, which is pre-set according to the cross-view feature matching requirements of protective clothing. The larger the value, the smoother the distribution of similarity between samples and the lower the distinction between positive and negative samples; the smaller the value, the higher the similarity distinction and the more significant the constraint effect. It is used to adjust the distribution scale of feature similarity and ensure the balance between the stability of loss calculation and the distinction.

[0059] The initial features of the target protective suit serve as the benchmark features for loss calculation. The accuracy of its representation directly determines the accuracy of similarity calculation and loss value. It is the core benchmark for feature mapping optimization and represents the original features of the target protective suit to be aligned.

[0060] The features of cross-view samples of the same target are derived from the adaptive initial features of positive samples in the comparison of positive and negative sample pairs. They belong to the features of different perspectives of the same protective suit and are used for positive constraints to improve the similarity between the initial features of the target protective suit and the features of cross-view samples of the same target. This directly promotes the spatial alignment of the features of cross-view samples of the same target and guides the network to learn the homology association rules of features of different perspectives of the same protective suit.

[0061] The features of heterogeneous targets are derived from the adaptive initial features of negative samples in the positive-negative sample pairs through comparative learning. These features belong to different protective suits and are used for negative constraints to reduce the similarity between the initial features of the target protective suit and the features of heterogeneous targets, increase the spatial distance between the features of heterogeneous targets, avoid the network from confusing the features of different protective suits, and strengthen the specific distinguishability of the features.

[0062] The cross-view sample association weight is extracted from the feature association parameter set. Based on the strength of the feature association between samples, the similarity between the initial features of the target protective clothing and the features of the cross-view samples of the same target is weighted and adjusted. The higher the association strength, the larger the weight value and the stronger the constraint effect. This is used to weight and strengthen the constraint effect of highly associated samples and improve the targeting of feature mapping optimization.

[0063] Cosine similarity function This is a function for calculating the similarity in the feature space. It is used to quantify the degree of similarity between feature vectors and is the core operational basis for loss calculation. It objectively measures the degree of correlation between the target feature and the positive and negative sample features.

[0064] Furthermore, the cross-view sample association weights are obtained through two core operations: feature matching degree calibration and viewpoint bias correction. The specific implementation steps are as follows: Extract core representation dimensions such as contour, texture, structural splicing position, and protective component shape of the target protective suit's initial features and cross-view sample features of the same target, and count the number of feature matching items in each dimension; The feature matching degree is taken as the ratio of the number of consistent feature matching items to the total number of feature dimensions. The matching degree is then normalized to obtain the initial association weight baseline value. Determine the degree of viewing angle deviation between the cross-view sample and the target sample, and set a corresponding weight correction coefficient based on the size of the viewing angle deviation. The smaller the viewing angle deviation, the larger the correction coefficient. The initial association weight baseline value is multiplied by the weight correction coefficient to obtain the final cross-view sample association weight, which is then included in the feature association parameter set for later use.

[0065] Based on the feature mapping constraint loss, the feature mapping weights of the contrastive learning feature extraction network are re-adjusted to gradually correct the weight parameters so that the feature mapping output fits the feature association requirements of positive and negative samples.

[0066] Furthermore, the weight retargeting adjustment specifically involves determining the correction direction of the feature mapping weights based on the numerical value and bias direction of the feature mapping constraint loss; iteratively correcting the feature mapping weights along the direction of loss reduction; verifying the degree of matching between the feature mapping output and the positive and negative sample feature association requirements after each correction; repeating the correction and verification until the feature mapping output meets the preset matching standard, thus completing the weight retargeting adjustment.

[0067] Based on the parameter status after the completion weight redirection adjustment, the feature mapping path of the contrastive learning feature extraction network is regularized and optimized, the transmission direction and processing logic of feature mapping are sorted out and standardized, and the optimized feature mapping process of protective clothing is obtained.

[0068] Furthermore, the optimized feature mapping process takes the adaptive initial features as input, completes the feature space dimension transformation according to the adjusted feature mapping weights, performs directional calibration of the feature distribution according to the normalized mapping path, eliminates the spatial distribution deviation of cross-view features of the same target, distinguishes the spatial distribution differences of features of different targets, and finally outputs cross-view aligned features with unified perspective and standardized representation.

[0069] According to the processing rules set in the optimized feature mapping process, directional feature mapping processing is carried out on the adaptive initial features of the protective clothing to unify the spatial distribution of features under different perspectives, eliminate feature distribution deviations caused by perspective differences, and generate cross-view alignment features of the protective clothing.

[0070] Furthermore, the optimized feature mapping process is configured with specific processing rules including cross-view feature space aggregation rules for the same target, feature space differentiation rules for different targets, weight adjustment and adaptation execution rules, and automatic correction rules for feature distribution deviations. These rules require the mapping process to bring the spatial distance between features from different viewpoints of the same protective suit closer and to widen the spatial distance between features from different protective suits. Based on the adjusted feature mapping weights, feature dimension transformation and distribution calibration are performed along a regular path to automatically eliminate cross-view feature space offsets and distribution disorder, ensuring the viewpoint uniformity and representation accuracy of the output features.

[0071] Cross-view alignment features include core features such as the overall outline of the protective suit, fabric texture, logo patterns, structural splicing positions, and the shape and size ratio of protective components, all of which are spatially calibrated. Each feature is unified to the same feature space dimension, eliminating spatial distribution deviations of cross-view features for the same target, preserving the unique characterization attributes of the protective suit itself, and possessing the ability to distinguish features from different targets. It can truly reflect the normalized appearance and structural features of the same protective suit from different perspectives.

[0072] By constructing positive and negative sample pairs for contrastive learning based on the viewpoint association attributes of cross-view protective clothing images, and using sample constraints as a guide to carry out feature mapping orientation constraints, weight retargeting adjustments, and mapping path regularization on the contrastive learning feature extraction network, the spatial distribution deviation of features under different viewpoints of the same protective clothing can be effectively eliminated, the distinguishability of features of different protective clothing can be enhanced, and cross-view alignment features with unified spatial dimensions and accurate representation can be generated. This provides standardized and highly recognizable feature data for subsequent feature processing, and significantly improves the stability and accuracy of cross-view protective clothing feature matching.

[0073] C4: Based on the twin encoder combined with the attention mechanism, the cross-view alignment features are enhanced to obtain the enhanced matching features of the protective suit; In this embodiment of the invention, the step of performing feature enhancement processing on the cross-view alignment features based on a twin encoder combined with an attention mechanism to obtain the enhanced matching features of the protective suit includes: The cross-view alignment features are input into the twin encoder, and the cross-view alignment features are synchronously parsed through the dual-branch structure of the twin encoder to obtain the dual-branch synchronous parsing features of the protective suit. Extract the representation dimension and distribution information of the cross-view alignment features, and construct an attention mechanism that adapts to the dual-branch synchronous parsing features based on the representation rules of the core features of the protective suit. Based on the attention mechanism, key feature filtering is performed on the dual-branch synchronous parsing features to locate the core representation information of the protective clothing in the dual-branch synchronous parsing features. Based on the core representation information, the dual-branch synchronous parsing features are weighted and integrated to obtain the enhanced matching features of the protective suit.

[0074] It should be specifically explained that the cross-view alignment features are input into the twin encoder, and the cross-view alignment features are parsed synchronously through the dual-branch synchronous processing architecture of the twin encoder to fully extract the spatial distribution and detailed representation information of the features, thus obtaining the dual-branch synchronous parsed features of the protective clothing.

[0075] Furthermore, the twin encoder adopts a dual-branch parallel processing architecture with consistent structure and shared parameters, which can perform synchronous and rule-based feature extraction and encoding parsing on the cross-view alignment features of the input. This ensures that the dual-branch output features have the same dimension, temporal synchronization, and consistent processing standards, and is used to generate dual-branch synchronous parsing features, providing a standardized feature foundation for the key feature selection and enhancement of the subsequent attention mechanism.

[0076] The dual-branch synchronous parsing feature is a dual-branch isomorphic architecture with shared parameters adopted by the twin encoder. It generates two sets of parallel homogeneous features after synchronous encoding and parsing of cross-view alignment features. The two sets of features have the same dimension, synchronous processing time, and consistent representation rules. They completely preserve the core representation information such as the outline of the protective clothing, fabric texture, structural splicing position, and protective component shape, without viewpoint bias or feature distortion. It provides standardized parallel feature data for key feature selection and enhancement of the attention mechanism.

[0077] Collect all representation dimensions and feature space distribution information of cross-view alignment features, and combine them with the inherent representation rules of core features such as the overall outline of the protective clothing, fabric texture, and protective components to build an attention mechanism that is compatible with the dual-branch synchronous parsing features.

[0078] By utilizing the constructed attention mechanism, key feature retrieval and screening are performed on the dual-branch synchronous parsing features to accurately locate the core representation information that reflects the ontological attributes of the protective suit in the dual-branch synchronous parsing features.

[0079] Furthermore, the core representation information is the key feature information that can uniquely identify the attributes of the protective clothing itself, which is selected from the dual-branch synchronous parsing features through the attention mechanism. Specifically, it includes the overall outline of the protective clothing, fabric texture, identification patterns, structural splicing positions, and the shape and size ratio features of protective components. This type of information has high recognizability, high stability and cross-view consistency, and is the core basis for feature enhancement and matching.

[0080] Based on the core representation information obtained from the positioning, the feature components corresponding to the core representation information in the dual-branch synchronous parsing features are weighted and enhanced. Then, the dual-branch features with enhanced weights are fused and integrated to obtain the enhanced matching features of the protective suit.

[0081] Furthermore, the specific steps of the weight enhancement process include locating the corresponding feature components and their distribution positions in the dual-branch synchronous parsing features of the core representation information; setting enhanced weights higher than those of ordinary feature components for the feature components corresponding to the core representation information; weighted coupling of the enhanced weights with the corresponding feature components to enhance the response intensity of the core features; and retaining the original weights of the non-core feature components to maintain the integrity of the overall feature structure.

[0082] The enhanced matching features are high-recognition standardized matching features of protective clothing generated after synchronous parsing by the twin encoder and weight enhancement integration by the attention mechanism. Specifically, they include core characterization information such as the overall outline of the protective clothing, fabric texture, logo patterns, structural splicing positions, shape and size ratio of protective components, etc., which are enhanced by weights. Redundant noise features are eliminated, while cross-view consistency and ontology-specific distinguishing attributes are retained. The feature response intensity is higher and the characterization is more prominent, which can be directly used for accurate cross-view matching of protective clothing.

[0083] By using a dual-branch isomorphic architecture of a twin encoder to achieve synchronous standardized parsing of cross-view alignment features, and combining an attention mechanism to accurately locate and enhance the core representation information of protective clothing and eliminate redundant feature interference, the feature recognition and response intensity can be significantly improved, generating highly stable and highly discriminative enhanced matching features, providing a high-quality feature foundation for subsequent accurate cross-view matching of protective clothing.

[0084] C5: Confirm the matching of the enhanced matching features and output the cross-view feature matching result of the protective suit.

[0085] In this embodiment of the invention, the step of confirming the matching of the enhanced matching features and outputting the cross-view feature matching result of the protective suit includes: The enhanced matching features are subjected to bidirectional feature matching verification, and the feature matching confidence is determined based on the bidirectional feature matching verification results. Based on the feature matching confidence, the enhanced matching features are effectively filtered to obtain the target matching features of the protective suit. Based on the target matching features, the cross-view feature matching results of the protective suit are generated and output.

[0086] It should be specifically noted that a two-way feature matching verification is performed on the enhanced matching features. The enhanced matching features to be matched are compared with the benchmark enhanced matching features in both forward and reverse feature comparisons to obtain the consistency results of the two-way feature matching.

[0087] Based on the consistency results obtained from the bidirectional feature matching verification, the feature matching confidence between the enhanced matching feature to be matched and the benchmark enhanced matching feature is determined, thereby determining the reliability of the feature matching.

[0088] Furthermore, the feature matching confidence score is a quantitative indicator of feature matching reliability generated based on the bidirectional feature matching verification results. It is used to characterize the degree of consistency between the enhanced matching features to be matched and the benchmark enhanced matching features in the core representation information of the protective clothing. The higher the value, the stronger the homology and consistency of the two sets of features and the more reliable the matching result, providing a core judgment basis for the effective screening of enhanced matching features.

[0089] Based on the feature matching confidence obtained from the determination, the enhanced matching features are effectively filtered, and features that meet the preset confidence requirements are retained to obtain the target matching features of the protective clothing.

[0090] Furthermore, the target matching feature is a highly reliable enhanced matching feature of the protective clothing that meets the pre-set confidence threshold after effective filtering by feature matching confidence. Specifically, it includes core characterization information such as the overall outline of the protective clothing, fabric texture, logo pattern, structural splicing position, shape and size ratio of protective components, which are weighted and meet the matching consistency standard. This feature eliminates low-confidence redundant features and has high homology, high consistency and high matching reliability, and can be directly used to generate accurate cross-view feature matching results.

[0091] Based on the target matching features obtained through screening, the cross-view feature matching results of the protective clothing are generated according to preset rules, and the matching results are output.

[0092] Furthermore, the preset rules are standardized judgment and output criteria for generating and outputting cross-perspective feature matching results based on target matching features. Specifically, these include feature homology judgment rules, confidence validity verification rules, matching result hierarchical labeling rules, and unified output format rules. These rules are used to determine whether cross-perspective features are homologous, verify matching validity, label matching levels, and standardize the output format, ensuring that the generated cross-perspective feature matching results are accurate, standardized, and have a unified judgment standard.

[0093] By employing a dual-branch isomorphic architecture of a twin encoder to achieve synchronous standardized parsing of cross-view alignment features, the consistency and standardization of feature processing are ensured. Combined with an attention mechanism, the core representation information of the protective clothing is accurately extracted and enhanced, and redundant feature interference is eliminated. This significantly improves the recognizability, response strength and stability of features, and generates highly accurate and discriminative enhanced matching features that are suitable for cross-view matching, providing high-quality and reliable feature support for subsequent feature matching stages.

[0094] like Figure 2 The diagram shown is a functional block diagram of a cross-view protective clothing feature matching system based on contrastive learning, provided by an embodiment of the present invention.

[0095] The cross-view protective clothing feature matching system 100 based on contrastive learning described in this invention can be installed in an electronic device. Depending on the functions implemented, the cross-view protective clothing feature matching system 100 may include a cross-view acquisition module 101, an adaptive feature extraction module 102, a spatial mapping optimization module 103, a feature enhancement processing module 104, and a feature matching confirmation module 105. The module described in this invention can also be referred to as a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, and are stored in the memory of the electronic device.

[0096] In this embodiment, the functions of each module / unit are as follows: The cross-view acquisition module 101 is used to perform cross-view acquisition processing on the protective clothing to obtain a standard cross-view protective clothing image, and input the standard cross-view protective clothing acquisition image into the contrast learning feature extraction network. The adaptive feature extraction module 102 is used to embed the viewpoint adaptive coding into the contrastive learning feature extraction network, perform distortion correction and feature extraction on the standard cross-view protective clothing acquired image, and obtain the adaptive initial features of the protective clothing. The spatial mapping optimization module 103 is used to construct a contrastive learning positive and negative sample pair based on the correlation relationship of the cross-view protective clothing images, and to perform spatial mapping optimization on the contrastive learning feature extraction network based on the contrastive learning positive and negative sample pair to obtain the cross-view alignment features of the protective clothing. The feature enhancement processing module 104 is used to perform feature enhancement processing on the cross-view alignment features based on the twin encoder combined with the attention mechanism to obtain the enhanced matching features of the protective suit. The feature matching confirmation module 105 is used to confirm the matching of the enhanced matching features and output the cross-view feature matching result of the protective clothing.

[0097] In the several embodiments provided by this invention, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0098] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0099] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0100] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0101] The embodiments of this application can acquire and process relevant data based on an artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0102] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A cross-view protective clothing feature matching method based on contrastive learning, characterized in that, The method includes: C1: Perform cross-view acquisition processing on the protective suit to obtain a standard cross-view protective suit image, and input the standard cross-view protective suit acquired image into the contrast learning feature extraction network; C2: Embed the viewpoint adaptive coding into the contrastive learning feature extraction network to perform distortion correction and feature extraction on the standard cross-view protective clothing images to obtain the adaptive initial features of the protective clothing; C3: Based on the correlation between the cross-view protective clothing images, construct contrastive learning positive and negative sample pairs, and based on the contrastive learning positive and negative sample pairs, perform spatial mapping optimization on the contrastive learning feature extraction network to obtain the cross-view alignment features of the protective clothing; C4: Based on the twin encoder combined with the attention mechanism, the cross-view alignment features are enhanced to obtain the enhanced matching features of the protective suit; C5: Confirm the matching of the enhanced matching features and output the cross-view feature matching result of the protective suit.

2. The cross-view protective clothing feature matching method based on contrastive learning as described in claim 1, characterized in that, The process of acquiring and processing cross-view images of protective suits to obtain standard cross-view images of protective suits, and then inputting these standard cross-view images into a contrastive learning feature extraction network, includes: The original images of the protective suit acquired from multiple perspectives are screened for viewpoint consistency, and image data that does not match the target acquisition perspective are removed to obtain the protective suit image set. Image quality normalization is performed on the set of protective clothing images to obtain standard cross-view protective clothing images of the protective clothing; Based on preset features, the standard cross-view protective clothing image is format adapted, and the format adaptation result is transmitted to the feature input port of the contrastive learning feature extraction network.

3. The cross-view protective clothing feature matching method based on contrastive learning as described in claim 1, characterized in that, The step of embedding the viewpoint adaptive coding into a contrastive learning feature extraction network to perform distortion correction and feature extraction on the standard cross-view protective clothing images, thereby obtaining the adaptive initial features of the protective clothing, includes: Based on the perspective distortion attribute features in the standard cross-view protective clothing image, a perspective distortion attribute feature set is constructed, and the perspective distortion attribute feature set is customized and encoded to obtain the perspective adaptive encoding of the protective clothing. The viewpoint adaptive coding is connected to the feature preprocessing node of the contrastive learning feature extraction network to complete the adaptation and fusion of the viewpoint adaptive coding and the contrastive learning feature extraction network. Based on the aforementioned viewpoint adaptive coding, viewpoint distortion recognition is performed on the standard cross-viewpoint protective clothing image to obtain the distortion feature localization result of the protective clothing; Based on the distortion feature localization results, distortion correction and feature extraction are performed on the standard cross-view protective clothing image to obtain the adaptive initial features of the protective clothing.

4. The cross-view protective clothing feature matching method based on contrastive learning as described in claim 3, characterized in that, The step of integrating the viewpoint adaptive coding into the feature preprocessing node of the contrastive learning feature extraction network to complete the adaptation and fusion of the viewpoint adaptive coding and the contrastive learning feature extraction network includes: Locate the feature preprocessing node of the contrastive learning feature extraction network to determine the access position of the viewpoint adaptive coding; Based on the access location, a data path connection is established between the viewpoint adaptive coding and the feature preprocessing node to create a transmission link between the viewpoint adaptive coding and the feature preprocessing node. Based on the transmission link, the viewpoint adaptive coding and the contrastive learning feature extraction network are adapted in a coordinated manner to achieve the adaptation and fusion of the viewpoint adaptive coding and the contrastive learning feature extraction network.

5. The cross-view protective clothing feature matching method based on contrastive learning as described in claim 1, characterized in that, Based on the correlation between the cross-view protective clothing images, a contrastive learning positive and negative sample pair is constructed. Based on the contrastive learning positive and negative sample pair, the spatial mapping of the contrastive learning feature extraction network is optimized to obtain the cross-view alignment features of the protective clothing, including: Based on the viewpoint association attributes of the cross-view protective clothing images, the same-target cross-view samples and different-target samples in the cross-view protective clothing images are distinguished, and the contrastive learning positive and negative sample pairs of the cross-view protective clothing images are constructed. Based on the contrastive learning positive and negative sample pairs, feature mapping orientation constraints are applied to the contrastive learning feature extraction network to obtain the optimized feature mapping process for the protective suit. Based on the optimized feature mapping process, directional feature mapping processing is performed on the adaptive initial features of the protective suit to generate cross-view alignment features of the protective suit.

6. The cross-view protective clothing feature matching method based on contrastive learning as described in claim 5, characterized in that, The process of applying feature mapping orientation constraints to the contrastive learning feature extraction network based on the contrastive learning positive and negative sample pairs to obtain the optimized feature mapping process for the protective suit includes: Based on the set of feature association parameters of the contrastive learning positive and negative sample pairs, the feature mapping constraint loss of the contrastive learning feature extraction network is calculated. Based on the feature mapping constraint loss, the feature mapping weights of the contrastive learning feature extraction network are re-adjusted, and the feature mapping path of the contrastive learning feature extraction network is normalized based on the re-adjustment, resulting in the optimized feature mapping process of the protective suit.

7. The cross-view protective clothing feature matching method based on contrastive learning as described in claim 6, characterized in that, The formula for calculating the feature mapping constraint loss is as follows: ； In the formula, This represents the feature mapping constraint loss of the protective suit. Indicates the temperature coefficient. This indicates the initial characteristics of the target protective suit. This represents the cross-view sample features of the protective suit targeting the same object. This indicates the heterogeneous target sample characteristics of the protective suit. This represents the cross-view sample association weight of the protective suit. The cosine similarity function is used. Let be the natural logarithm function with base e. It is an exponential function with the natural constant e as its base.

8. The cross-view protective clothing feature matching method based on contrastive learning as described in claim 1, characterized in that, The method of enhancing the cross-view alignment features based on a twin encoder combined with an attention mechanism to obtain the enhanced matching features of the protective suit includes: The cross-view alignment features are input into the twin encoder, and the cross-view alignment features are synchronously parsed through the dual-branch structure of the twin encoder to obtain the dual-branch synchronous parsing features of the protective suit. Extract the representation dimension and distribution information of the cross-view alignment features, and construct an attention mechanism that adapts to the dual-branch synchronous parsing features based on the representation rules of the core features of the protective suit. Based on the attention mechanism, key feature filtering is performed on the dual-branch synchronous parsing features to locate the core representation information of the protective clothing in the dual-branch synchronous parsing features. Based on the core representation information, the dual-branch synchronous parsing features are weighted and integrated to obtain the enhanced matching features of the protective suit.

9. The cross-view protective clothing feature matching method based on contrastive learning as described in claim 1, characterized in that, The step of verifying the enhanced matching features and outputting the cross-view feature matching result of the protective suit includes: The enhanced matching features are subjected to bidirectional feature matching verification, and the feature matching confidence is determined based on the bidirectional feature matching verification results. Based on the feature matching confidence, the enhanced matching features are effectively filtered to obtain the target matching features of the protective suit. Based on the target matching features, the cross-view feature matching results of the protective suit are generated and output.

10. A cross-view protective clothing feature matching system based on contrastive learning, characterized in that, To implement the cross-view protective clothing feature matching method based on contrastive learning as described in claim 1, the system includes a cross-view acquisition module, an adaptive feature extraction module, a spatial mapping optimization module, a feature enhancement processing module, and a feature matching confirmation module, wherein: The cross-view acquisition module is used to perform cross-view acquisition processing on the protective suit to obtain a standard cross-view protective suit image, and input the standard cross-view protective suit acquisition image into the contrast learning feature extraction network. The adaptive feature extraction module is used to embed the viewpoint adaptive coding into the contrastive learning feature extraction network, perform distortion correction and feature extraction on the standard cross-view protective clothing images, and obtain the adaptive initial features of the protective clothing. The spatial mapping optimization module is used to construct contrastive learning positive and negative sample pairs based on the correlation relationship of the cross-view protective clothing images, and to perform spatial mapping optimization on the contrastive learning feature extraction network based on the contrastive learning positive and negative sample pairs to obtain the cross-view alignment features of the protective clothing. The feature enhancement processing module is used to perform feature enhancement processing on the cross-view alignment features based on the twin encoder combined with the attention mechanism to obtain the enhanced matching features of the protective suit. The feature matching confirmation module is used to confirm the matching of the enhanced matching features and output the cross-view feature matching results of the protective suit.

Images