Target recognition method, device and medium
By combining the knowledge base and data simulation module of the intelligent reconnaissance system with the XGBOOST model, the problems of low accuracy and difficulty in adapting to complex scenarios in traditional target recognition methods are solved, and high-precision multi-method joint target recognition is achieved.
Patent Information
- Application Number
- CN202310599297.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-05-25
- Publication Date
- 2025-10-24
- Estimated Expiration
- 2043-05-25
AI Technical Summary
Traditional target recognition methods have low accuracy and poor matching rule flexibility. Multi-classification methods rely on a large number of labeled samples and data features are difficult to align, making them unsuitable for complex real-world reconnaissance scenarios.
An intelligent reconnaissance system is adopted, which combines a knowledge base module, a data simulation module, and a recognition module. The system uses a knowledge graph to represent the detection knowledge, generates simulation data and performs matching feature extraction, and uses the XGBOOST ensemble learning model for recognition. This transforms a multi-classification problem into a binary classification problem, and uses recognition templates and similarity features to identify the target.
It improves recognition accuracy, reduces training sample requirements, adapts to complex target radiation source deployment scenarios, and enhances recognition performance and flexibility.
Smart Images
Figure CN116720102B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of target recognition technology, and in particular relates to a target recognition method, device and medium. Background Art
[0002] The intelligent reconnaissance system utilizes multiple front-end signal detection methods, such as ultra-short wave (VHF) and radar, to acquire various radiator signals related to the target. Leveraging existing digital signal processing technologies, it can identify existing radiators and further obtain measurement parameters of the radiator targets. The specific radiator parameters of these targets, such as the nominal radar RF value and VHF communication modulation method, as well as specific combinations of these radiator parameters, can serve as effective basis for target identification.
[0003] Traditional target recognition generally uses fixed matching rules and thresholds to identify targets that meet the rules. This method generally has low recognition accuracy and poor flexibility in setting matching rules. On the other hand, general intelligent recognition methods use supervised learning methods based on multi-classification. Their model training relies on a large number of labeled samples and requires data feature alignment. Multi-classification methods require sufficient data samples for each individual class, but in actual scenarios, the difficulty of obtaining samples of various types of targets varies. In particular, samples of important targets are often difficult to obtain. In addition, the situation of target-mounted radiation sources in real scenarios is complex. For example, the same means may have multiple devices, making data features difficult to align. Therefore, multi-classification methods are difficult to apply to real detection scenarios. Summary of the Invention
[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a target recognition method, device and medium, so as to realize multi-means joint target recognition with high recognition accuracy and easy training based on an intelligent reconnaissance system.
[0005] The object of the present invention is achieved through the following technical solutions:
[0006] A target recognition method is implemented using an intelligent detection system. The intelligent detection system includes a knowledge base module, a data simulation module, a preprocessing module, and a recognition module. The knowledge base module stores detection knowledge required for target recognition and constructs recognition templates for recognition tasks. The data simulation module generates simulated data based on the recognition templates for machine learning model training. The preprocessing module generates matching features required for recognition based on the recognition templates and the recognition data. The recognition module receives matching features from simulated or real data to implement recognition model training and target prediction.
[0007] The method specifically includes:
[0008] In response to the received target identification task, detect knowledge in the knowledge base module is queried to build an identification template for the target identification task;
[0009] The data simulation module selects an identification template, simulates the discovery of a target radiation source in an actual reconnaissance scene, and generates simulation detection data with noise according to the selected identification template;
[0010] The preprocessing module builds the target identification task and the identification template into a task-template matching pair, and extracts similarity-based matching features.
[0011] The identification module receives matching features of simulation or real data and sample labels, trains a rule-based identification model, and the identification model receives matching features and returns identification model discrimination results.
[0012] Further, the detection knowledge is represented by a knowledge graph and the knowledge graph is stored in the form of triples, the target entity and the attribute are regarded as the vertices of the directed graph, the relationship between entities is regarded as the edge of the directed graph, and the increase, deletion, query and modification of the knowledge base are realized by using the storage and search of the graph, the knowledge base module takes the vertex of the knowledge graph as a query target, and the query is completed by searching adjacent vertices and edges.
[0013] Further, the identification template is a query result of the knowledge base module, and the identification template is used to represent detection information of the target, and the detection information includes single value, multiple value and range value.
[0014] Further, the identification template is represented in the form of a tree.
[0015] Further, the system further comprises an online interface, and the method defines an external interface of the intelligent reconnaissance system target identification through the online interface and provides query services to the outside.
[0016] Further, the preprocessing module builds the target identification task and the identification template into a task-template matching pair, and extracts similarity-based matching features specifically includes:
[0017] A block algorithm is used to convert a multi-classification identification problem into a binary classification matching problem, n multi-classification identification tasks are combined with m matching templates to build n x m binary classification matching tasks.
[0018] Features of the n x m binary classification matching tasks are calculated by a matching feature extraction algorithm, and the features of the matching tasks are generated based on the similarity between the identification task data and the matching template.
[0019] Further, the rule-based identification model includes an XGBOOST-based ensemble learning model.
[0020] In another aspect, the present application also provides a computer device, which comprises a processor and a memory, and the memory stores a computer program, and the computer program is loaded and executed by the processor to implement any one of the above target identification methods.
[0021] In another aspect, the present application also provides a computer readable storage medium, which stores a computer program, and the computer program is loaded and executed by a processor to implement any one of the above target identification methods.
[0022] The present application has the following beneficial effects:
[0023] (1) The present application is directed to multi-method joint target identification of intelligent reconnaissance systems, and the method uses the detection results of multiple methods in the reconnaissance field to perform joint identification, and the identification effect is significantly improved compared with single-method methods. At the same time, the method effectively combines the advantages of traditional knowledge base identification and intelligent identification methods, and the method not only learns the knowledge of the discrimination model from the labeled sample data, but also fully considers the influence of knowledge base knowledge in the process of extracting features by matching feature algorithms, and fully utilizes the knowledge of multiple ways to improve the identification accuracy.
[0024] (2) The present application is directed to multi-method joint target identification of intelligent reconnaissance systems, and the characteristic is that the algorithm converts the multi-classification problem into a binary classification problem, so that the original multi-classification problem is converted from modeling according to data features to modeling according to the similarity between data features and identification templates, so that the learning of any target sample can affect other targets, thereby effectively reducing the demand for training samples and reducing the training difficulty.
[0025] (3) The present application is directed to multi-method joint target identification of intelligent reconnaissance systems, and the characteristic is that the identification template can describe the complex target radiation source carrying situation, including multiple methods, multiple devices, etc. The matching feature extraction algorithm fully considers the diversity of the identification template, can adapt to similarity feature calculation of different knowledge such as single value, multiple value, range value, etc., so that the model will not be limited by the alignment conditions of the features to reduce the accuracy. BRIEF DESCRIPTION OF DRAWINGS
[0026] Figure 1 is a schematic diagram of the target identification method system provided by the embodiment of the present application;
[0027] Figure 2 is a schematic diagram of the XGBOOST training algorithm flow provided by the embodiment of the present application. DETAILED DESCRIPTION
[0028] Following make the embodiments of the present application through specific, the person skilled in the art can easily understand the other advantages and efficacy of the present application from the disclosure of this specification. The present application can also be implemented or applied by another different embodiment, and the details in the specification can be based on different views and applications, and various modifications or changes can be made without departing from the spirit of the present application. It should be noted that the following examples and features in the examples can be combined with each other without conflict.
[0029] All other embodiments obtained by the person skilled in the art based on the embodiments in the present application without making creative efforts belong to the scope of protection of the present application.
[0030] The conventional target recognition generally finds the target meeting the rule by using the threshold value through the fixed matching rule. The method usually has low recognition accuracy and poor flexibility in setting the matching rule. On the other hand, the general intelligent recognition method adopts a supervised learning method based on multi-classification, and the model training thereof depends on a large number of labeled samples and requires data feature alignment. The multi-classification method needs to provide sufficient data samples for each individual class, and the acquisition of various types of target samples in the actual scene is difficult, especially the samples of important targets are often difficult to obtain. In addition, the target-mounted radiation source in the real scene is complex, for example, the same means may have multiple devices, and the data features are difficult to align, therefore, the multi-classification method is difficult to apply to the real detection scene.
[0031] In order to solve the above technical problems, the following embodiments of the target recognition method, device and medium of the present application are proposed.
[0032] Embodiment 1
[0033] Reference Figure 1 As Figure 1 shown is a schematic diagram of the system composition of the target recognition method provided by the present embodiment, which is realized based on an intelligent reconnaissance system.
[0034] In the intelligent reconnaissance system, the knowledge base module represents the detection knowledge of various means of the target by using the knowledge graph, and stores the knowledge graph in the form of a triple of "relation-object 1-object 2". The target knowledge base and the recognition template are used to describe the detection knowledge of various means of the target. The attributes of the target and the radiation source in the real scene are complex and difficult to describe in a unified format of traditional relational data. The knowledge base uses the knowledge graph technology to regard the target entity and the attribute as the vertex V of the directed graph, and the relationship between the entities as the edge E of the directed graph, and uses the storage, search and other technologies of the graph to realize the functions of adding, deleting, querying and modifying the knowledge base.
[0035] The recognition template construction algorithm constructs a recognition template for a target recognition task by querying a knowledge base target. The knowledge base query takes a platform, i.e., a vertex of a knowledge graph, as a query target, and completes the query by searching adjacent vertices and edges. The query result is represented in the form of a tree, referred to as a recognition template. The recognition template can represent complex detection information of a target in multiple ways, such as single value, multiple values, range values, and the like.
[0036] The data simulation module uses a random selection algorithm to select the recognition template, simulating the discovery of a target radiation source in an actual reconnaissance scene.
[0037] The data generation algorithm generates simulated detection data with noise according to the selected recognition template; the online interface defines an external interface for target recognition of the intelligent reconnaissance system, and implements query services at the application level.
[0038] The data preprocessing module uses a block algorithm to construct a "task-template" matching pair from the recognition task data and the recognition template, and extracts matching features based on similarity using a matching feature extraction algorithm; the recognition model receives matching features and sample labels of simulated or real data in the algorithm training stage, and trains the recognition model based on the XGBOOST method. The recognition model receives matching features of online queries in the algorithm prediction stage, and returns the recognition model discrimination result.
[0039] Let a given target set be T = {t1, t2,..., t n}, and let an entity set E = {e1, e2,..., e m} in the knowledge base, as shown in Table 1 and Table 2, respectively.
[0040] Objectives Attribute 1 Attribute 2 … Identification tag t1 60 90 … e1 t2 70 … t3 80 …
[0041] Table 1: Target Set Table
[0042] Entity Attribute 1 Attribute 2 … e1 63 92 … e2 90 23 … e3 103 29 …
[0043] Table 2: Entity Set
[0044] According to the target set and the entity set, n x m matching pairs are generated, denoted as D. The labeled data in D is denoted as D *The pairs with the correct corresponding entities in the library are generated with a label of 1, and the rest of the entities are generated with a label of 0 for mismatched pairs. Similarity is generated, first analyzing the characteristics of the input attributes: the output data is mainly the processing result obtained by the front signal recognition, containing signal, radar and other measurement characteristics rather than the signal itself. These data are mainly real numbers, nominal quantities. Therefore, the processing method is also mainly divided into two categories: for single-value nominal quantities, such as frequency hopping mode, there is no similarity between these data, only considering whether they match, the main feature is to directly match the knowledge base, taking 1 for matched data and 0 for unmatched data. If the instances in the knowledge base accept multiple values, any match takes 1.
[0045] The matching pairs after blocking are shown in Table 3.
[0046]
[0047]
[0048] Table 3 Matching pair table
[0049] For multi-value nominal quantities, that is, the case where multiple values can exist at the same time, JACCARD similarity measure can be used for calculation.
[0050] JACCRD(A,B) = (|A∩B|) / (|A∪B|)
[0051] For continuous data, such as maximum, minimum, and average values of frequency, the similarity of these values is calculated by different measurement methods, such as percentage and absolute difference.
[0052] The data preprocessing module uses a blocking algorithm to convert the multi-classification recognition problem into a binary classification matching problem, achieving effective training under small sample conditions. The binary classification recognition problem can effectively utilize existing knowledge base knowledge and reduce the training difficulty.
[0053] Reference Figure 2 As Figure 2 shown is a schematic diagram of the XGBOOST training algorithm flow of the embodiment. XGBOOST is an ensemble learning algorithm, and its idea is to integrate multiple CART decision trees to make a decision together to improve prediction accuracy. XGBOOST is based on the BOOST algorithm and generates multiple CART decision trees through iterative methods. In the s-th iteration, the objective function of the s-th tree is the residual of the first s-1 rounds of objective functions.
[0054] In each iteration, the algorithm first calculates the first-order gradient g i and the second-order gradient h i of each sample point in the objective function.
[0055]
[0056] Then, XGBOOST trains the CART tree of the current round using the training data. During the training process, CART is generated in a recursive manner, and each time the CART tree determines the best attribute and split point according to the evaluation target of formula 2:
[0057]
[0058] where G and H represent the sum of the first-order and second-order gradients of all target functions in a sample set I, respectively, that is:
[0059] G I =∑ i∈I g i H I =∑ i∈I h i
[0060] λ is an L2 regularization parameter used to control the complexity of the model. The stopping condition of the split is usually that the current data purity reaches a threshold, the split layer reaches an upper limit, or the current data set is less than a threshold. For the generated leaf node, the model calculates the weight of the leaf according to the following formula:
[0061]
[0062] After recursive generation, the finally generated tree f s (x i ) is added to the previous model:
[0063]
[0064] where η represents the learning rate.
[0065] According to the actual reconnaissance scene, the feature dimension of the target recognition task is small, which is embodied in the traditional real number and nominal data in the data set, and there are sparse features. XGBOOST is based on CART tree, so the model is a rule-based model, which can intuitively explain the judgment standard.
[0066] The specific process of the target recognition method provided in the embodiment is as follows:
[0067] In response to the received target recognition task, the detection knowledge in the knowledge base module is queried to construct a recognition template for the target recognition task;
[0068] The data simulation module selects the recognition template, simulates the discovery situation of the target radiation source in the actual reconnaissance scene, and generates simulation detection data with noise according to the selected recognition template;
[0069] The preprocessing module constructs the target recognition task and the recognition template into a task-template matching pair, and extracts matching features based on similarity;
[0070] The recognition module receives the matching features of the simulation or real data and the sample label, trains a rule-based recognition model, the recognition model receives the matching features, and returns a recognition model discrimination result.
[0071] The embodiment is directed to multi-means joint target recognition of an intelligent reconnaissance system. The method uses detection results of multiple means in the reconnaissance field for joint recognition, and the recognition effect is significantly improved compared with a single means method. Meanwhile, the method effectively combines the advantages of traditional knowledge base recognition and intelligent recognition methods. The method not only learns the knowledge of the discrimination model from the labeled sample data, but also fully considers the influence of knowledge base knowledge in the process of extracting features, and fully utilizes multiple ways of knowledge to improve the recognition accuracy.
[0072] The embodiment is directed to multi-means joint target recognition of an intelligent reconnaissance system. The characteristic is that the algorithm converts a multi-classification problem into a binary classification problem, so that the original multi-classification problem is converted from modeling for data features into modeling for the similarity between data features and recognition templates. Learning for any target sample can affect other targets, thereby effectively reducing the demand for training samples and reducing the training difficulty.
[0073] The embodiment is directed to multi-means joint target recognition of an intelligent reconnaissance system. The characteristic is that the recognition template can describe a complex target radiation source carrying condition, including multiple means, multiple devices, etc. The matching feature extraction algorithm fully considers the diversity of the recognition template, can adapt to similarity feature calculation of different knowledge such as single value, multiple value, and range value, so that the model will not be affected by the feature alignment restriction condition to reduce the accuracy.
[0074] Embodiment 2
[0075] The preferred embodiment provides a computer device that can implement the steps in any embodiment of the target recognition method provided in the embodiments of the present application, and thus can achieve the beneficial effects of the target recognition method provided in the embodiments of the present application. For details, see the previous embodiments, which will not be repeated here.
[0076] Embodiment 3
[0077] Those skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructions, or controlled by related hardware by instructions. The instructions can be stored in a computer readable storage medium and loaded and executed by a processor. Therefore, the embodiment of the present application provides a storage medium, which stores a plurality of instructions. The instructions can be loaded by a processor to execute the steps in any embodiment of the target recognition method provided in the embodiments of the present application.
[0078] The storage medium can include a read-only memory (ROM), a random access memory (RAM), a magnetic disk, an optical disk, or the like.
[0079] Due to the instructions stored in the storage medium, the steps in any target recognition method embodiment provided by the embodiments of the present application can be executed, thus achieving the beneficial effects of any target recognition method provided by the embodiments of the present application. Details are described in the foregoing embodiments, which will not be repeated here.
[0080] The above merely describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, and improvement within the spirit and principle of the present application shall be included in the protection scope of the present application.
Claims
1. A target recognition method characterized by, The method is realized by an intelligent investigation system, the intelligent investigation system comprises a knowledge base module, a data simulation module, a preprocessing module and an identification module, the knowledge base module stores detection knowledge required for target identification and constructs an identification template for an identification task, the data simulation module generates simulation data according to the identification template for machine learning model training, the preprocessing module generates matching features required for identification according to the identification template and identification data, and the identification module receives matching features of simulation or real data to realize identification model training and target prediction. The method specifically comprises: In response to a received target identification task, detection knowledge in the knowledge base module is queried to construct an identification template for the target identification task; The data simulation module selects the identification template, simulates the discovery of a target radiation source in an actual investigation scene, and generates simulation detection data with noise according to the selected identification template; The preprocessing module constructs a task-template matching pair of the target identification task and the identification template, and extracts matching features based on similarity; The identification module receives matching features of simulation or real data and sample labels, trains a rule-based identification model, the identification model receives matching features, and returns identification model discrimination results.
2. The target identification method of claim 1, wherein, The detection knowledge is represented by a knowledge graph, and the knowledge graph is stored in the form of triples, a target entity and an attribute are regarded as vertices of a directed graph, and a relationship between entities is regarded as an edge of the directed graph, the increase, deletion, query and modification of the knowledge base are realized by using graph storage and search, the knowledge base module takes vertices of the knowledge graph as query targets, and queries are completed by searching adjacent vertices and edges.
3. The object recognition method of claim 2, wherein, The identification template is a query result of the knowledge base module, the identification template is used to represent detection information of a target, and the detection information comprises single values, multiple values and range values.
4. The object recognition method of claim 3, wherein, The identification template is represented in the form of a tree.
5. The object recognition method of claim 1, wherein, The system further comprises an online interface, and the method defines an external interface of target identification of the intelligent investigation system through the online interface, and provides query services to the outside.
6. The target identification method of claim 1, wherein, The preprocessing module constructs a task-template matching pair of the target identification task and the identification template, and extracts matching features based on similarity, specifically comprising: A block algorithm is used to convert a multi-classification identification problem into a binary classification matching problem, n multi-classification identification tasks are combined with m matching templates to construct n*m binary classification matching tasks; Features of the n*m binary classification matching tasks are calculated by a matching feature extraction algorithm, and the features of the matching tasks are generated based on similarity between identification task data and matching templates.
7. The target identification method of claim 1, wherein, The rule-based identification model comprises an integrated learning model based on XGBOOST.
8. A computer device, comprising: The computer device comprises a processor and a memory, the memory stores a computer program, the computer program is loaded and executed by the processor to realize the target identification method in any one of claims 1-7.
9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, the computer program is loaded and executed by the processor to realize the target identification method in any one of claims 1-7.