A multi-agent credible collaborative target recognition method based on uncertainty quantification in an open environment

Abstract

The embodiment of the present invention discloses a multi-agent trusted collaborative target recognition method based on uncertainty quantification in an open environment, which includes: S1, creating a distributed communication network for multiple agents in an open environment, and each agent is independently loaded with a corresponding target recognition model; S2, based on the uncertainty quantification mechanism, target recognition is performed on the task to be recognized through each target recognition model, and the corresponding target recognition results are obtained; S3, based on the multi-agent consistency mechanism, through the communication path provided by the distributed communication network, each agent is controlled to perform the first global information exchange, so as to exclude the exceptional data in the task to be recognized through the first global information exchange; S4, based on the multi-agent consistency mechanism, through the communication path provided by the distributed communication network, each agent is controlled to perform the second global information exchange, so as to obtain the final recognition prediction result through the second global information exchange. The present invention introduces the uncertainty quantification mechanism to enable the present invention to distinguish the out-of-distribution samples in the open scene; at the same time, by introducing the multi-agent consistency mechanism, the global information exchange is performed between the models, and the credibility of the collaborative recognition decision is guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of multi-agent collaborative technology, and in particular to a multi-agent trusted collaborative target recognition method based on uncertainty quantification in an open environment. Background Art

[0002] As a powerful machine learning model, neural networks have achieved great success in image recognition, natural language processing, speech recognition and other fields. The core advantage of neural networks lies in their powerful feature extraction and pattern recognition capabilities, which makes them excellent in the field of target recognition. With the improvement of computing power and the increase of data volume, neural network models have become deeper and more complex, and can capture more subtle and abstract data features. At present, many neural network models are based on a single network hypothesis. However, in the complex environment of the real world, a single model is often difficult to cope with all situations independently. This requires multiple models to work together to improve the robustness and accuracy of the overall system.

[0003] Target collaborative recognition is one of the typical multi-model collaborative tasks. Compared with using only a single model to complete target recognition, the collaborative recognition of multiple models is conducive to improving the robustness and fault tolerance of recognition, thereby ensuring the adaptability of the model to different tasks. Target collaborative recognition has been widely used in many fields in reality. For example, in the field of national defense security and military reconnaissance, with the trend of miniaturization, agility, and concealment of targets, traditional single platforms have problems such as poor accuracy and difficulty in recognition due to factors such as fixed reconnaissance direction, single angle, limited reconnaissance capability, and weak anti-interference and anti-destruction capabilities. Therefore, target collaborative recognition technology is needed to support multiple intelligent platforms to collaboratively identify targets from different areas and directions in a distributed manner and reach a decision-making consensus.

[0004] The past decade has witnessed the rapid development of multi-agent collaboration technology. Multi-agent collaboration technology supports the interaction and collaboration between multiple agents, thus forming multiple agents into an "invisible" whole to complete complex tasks. At present, a considerable amount of research results have been accumulated in the field of multi-agent collaboration. Dong W proposed a multi-agent-based information collaboration computing framework. It discusses the process structure of information collaboration computing and analyzes the process of multi-agent information collaboration tasks according to the task life cycle. Li studied the performance of multi-agent reinforcement learning algorithms in noisy environments and proposed a model called multi-agent actor critic collaboration (MACC) to deal with noise interference. In response to bandwidth limitations and channel noise problems, He proposed task-oriented communication principles in multi-agent collaborative communication. At the same time, in recent years, target collaborative recognition technology has received widespread attention from researchers. According to different task types, target collaborative recognition tasks are generally divided into open set recognition tasks and closed set recognition tasks. Su H proposed an improved classifier and ensemble learning strategy to solve the difficult problem of wetland feature hyperspectral image classification. Hua G proposed a multi-expert collaborative active learning model. It can handle label inputs with different noise levels and detect irresponsible labelers online. Bu L proposed a food image recognition method based on transfer learning and ensemble learning to improve the recognition accuracy. Different basic learner combination strategies were adopted to build an ensemble model to classify feature information.

[0005] However, the existing target collaborative recognition technology is not yet mature and can hardly meet the actual needs. It is mainly manifested in the following aspects. First, the intelligence level of multi-model decision fusion is low and the credibility is not strong. For example, the fusion method based on fixed weights does not consider the reliability of each decision between different tasks and different models. The decisions of different models under different tasks cannot be fused together through reasonable weights; second, it lacks the ability to handle open scenes. Most of the existing collaborative recognition frameworks and recognition models are based on the idea of ​​closed set recognition, but the real world is open, and the recognition task will face many targets that do not belong to the training set. These out-of-distribution targets are often more valuable in certain security fields. However, most of the existing target collaborative recognition technology frameworks and models cannot handle these "out-of-distribution" samples, but mistakenly identify them as a certain type of known samples, resulting in untrustworthy decision-making behavior. The cost of this behavior in the security field is unbearable. Summary of the invention

[0006] Based on this, in order to solve the shortcomings of the existing technology, a multi-agent trusted collaborative target recognition method based on uncertainty quantification in an open environment is proposed.

[0007] A multi-agent credible collaborative target recognition method based on uncertainty quantification in an open environment, characterized by comprising:

[0008] S1. Create a distributed communication network for multiple agents in an open environment, and each agent is independently loaded with its own corresponding target recognition model;

[0009] S2. Based on the uncertainty quantification mechanism, target recognition is performed on the task to be recognized through each target recognition model, and the corresponding target recognition results are obtained, wherein the target recognition results at least include: target recognition prediction data and uncertainty data, wherein the uncertainty data is used to characterize the uncertainty degree of the model for the target recognition prediction data, i.e., the reliability;

[0010] S3. Based on the multi-agent consistency mechanism, each agent is controlled to perform a first global information exchange through the communication path provided by the distributed communication network, so as to exclude the exception data in the task to be identified through the first global information exchange, wherein the information participating in the first global interaction is the uncertainty data;

[0011] S4. Based on the multi-agent consistency mechanism, each agent is controlled to perform a second global information exchange through the communication path provided by the distributed communication network, so as to obtain the final recognition prediction result through the second global information exchange, wherein the information participating in the second global interaction is the target recognition prediction data.

[0012] Optionally, in one of the embodiments, in S2, the posterior distribution form of the network parameters of the target recognition model is changed through an uncertainty quantification mechanism, namely, the uncertainty quantification method MC-dropout, and multiple detections are performed on the same input data to obtain target recognition results of each target recognition model for the task to be recognized, and the target recognition results include at least: target recognition prediction data and uncertainty data, and the uncertainty data is used to characterize the uncertainty degree of the model for the target recognition prediction data, i.e., the reliability; the target recognition prediction data is the probability distribution data corresponding to the target recognition model for different recognition categories.

[0013] Optionally, in one embodiment, S2 further provides the following assumptions, which include:

[0014] Assume that each recognition model M i , the recognition model M i The recognition ability is to be able to identify q-type targets, and its corresponding input data x i , identify the network parameters as Then for the input data x i , test H times, each test corresponds to the output data is a (1x q) vector, and the corresponding sign function is h∈(1, 2, ..., H), the final target recognition result is the target recognition prediction data and uncertainty data. The target recognition prediction data is the i The predicted class probability distribution E(y i ); uncertainty data Var(y i ) is the prediction variance, which represents the uncertainty measure for the inference; through the above uncertainty quantification mechanism, the probability distribution corresponding to each recognition category and the result corresponding to the uncertainty can be obtained, which is expressed as c i ={E(y i ),Var(y i )}.

[0015] Optionally, in one embodiment, the specific steps of S3 include:

[0016] S31, setting global information exchange constraints and initializing;

[0017] The setting of global information exchange constraints specifically refers to setting the communication starting node t0 and the number of times n that global communication is completed between nodes, thereby controlling the time of information exchange between multiple agents;

[0018] Initialization specifically refers to setting and initializing an array A for storing the state solutions of each multi-agent at each communication time point;

[0019] S32, forming a communication feature vector corresponding to each intelligent agent, and controlling each intelligent agent to perform the first global information exchange through the communication path provided by the distributed communication network, that is, sending the local communication feature vector to the neighboring node and receiving the communication feature vector sent by the neighboring node, thereby obtaining the uncertainty fusion result after the first global information exchange In the first global information exchange, each agent exchanges information with its corresponding neighboring nodes n times, and in each information exchange process, a consensus algorithm is used to calculate the communication results of each agent after each exchange, i.e., the uncertainty;

[0020] The communication feature vector is used to characterize the first initial state corresponding to each agent, and the element of the feature vector is the target recognition model M i The preliminary prediction results obtained i The uncertainty in ; that is, the uncertainty is set as a parameter that characterizes the initial state of the agent;

[0021] After forming the communication feature vector corresponding to each intelligent agent, each intelligent agent performs n information exchanges or information exchanges with its corresponding neighboring nodes through the communication path provided by the distributed communication network, so that in each exchange, the local communication feature vector is sent to the neighboring node and the communication feature vector sent by the neighboring node is received, and finally the uncertainty fusion result corresponding to each recognition category after n information exchanges is obtained.

[0022] S33, based on the given threshold value, the uncertainty fusion result Screening is performed to exclude exceptional data, namely OOD data, in the tasks to be identified.

[0023] Optionally, in one embodiment, the specific steps of S4 include:

[0024] S41, forming a communication feature matrix corresponding to each intelligent agent, and controlling each intelligent agent to perform a second global information exchange through the communication path provided by the distributed communication network, that is, sending the local communication feature matrix to the neighboring node and receiving the communication feature matrix sent by the neighboring node, and then obtaining the final prediction fusion result after the second global information exchange, that is, the fusion result E(y of the prediction class probability distribution) i ); In the second global information exchange, each agent exchanges information with its corresponding neighboring nodes n times, and in each information exchange process, a consensus algorithm is used to calculate the communication results of each agent after each exchange, that is, the probability distribution value;

[0025] The AC feature matrix is ​​used to characterize the second initial state corresponding to each agent. In the elements of the AC feature matrix, each row corresponds to the test set-screened input data x i The classification result of each image sample in is the target recognition model M i The respective prediction results c i The probability values ​​in each column correspond to the category specified by the data set; then after forming the communication feature matrix corresponding to each intelligent agent, each intelligent agent exchanges information with its corresponding neighboring node n times through the communication path provided by the distributed communication network, so that in each exchange, the local communication feature matrix is ​​sent to the neighboring node and the communication feature matrix sent by the neighboring node is received, and finally the prediction fusion result corresponding to each recognition category after n information exchanges is obtained;

[0026] S42, the prediction fusion results corresponding to each identification category after the obtained n information exchanges are input into the final classification layer, and the prediction labels of the data are obtained to complete the information exchange between multiple models, and then the final c i .

[0027] Optionally, in one of the embodiments, in S1, a distributed communication network is formed by obtaining a directed graph corresponding to each intelligent agent and obtaining an adjacency matrix corresponding to the directed graph G.

[0028] In addition, a computer-readable storage medium is also proposed, comprising computer instructions. When the computer instructions are executed on a computer, the computer is enabled to execute the method.

[0029] Implementing the embodiments of the present invention will have the following beneficial effects:

[0030] The present invention proposes a multi-agent trusted collaborative target recognition method (UQ-MCTR) based on uncertainty quantification in an open environment, and designs a decision fusion framework for multi-agent distributed conditions suitable for collaborative recognition tasks. The distributed multi-agents supporting the open environment achieve decision consistency for the target recognition task through the framework, thereby ensuring the credibility of the collaborative recognition decision. The uncertainty quantification method is introduced into the framework to perform uncertainty assessment for each decision of each recognition model. The uncertainty is sensitive to the specific recognition task, and the uncertainty value reflects the "confidence" of the specific model in the specific decision. Then, by designing a reasonable information consistency transmission mechanism, the decisions of each agent are fused to realize automatic recognition of out-of-distribution targets that are prone to appear in an open environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0032] in:

[0033] Figure 1 The following is a flowchart of the basic steps corresponding to the method of the present invention;

[0034] Figure 2 It is an implementation scenario example diagram, an algorithm flow example diagram, and a data exchange example diagram in one embodiment of the present invention;

[0035] Figure 3 This is a picture example diagram of part of the CIFAR-10 dataset described in the present invention;

[0036] Figure 4 The performance curve of the method described in the present invention under different uncertainties;

[0037] Figure 5The performance curve of the method described in the present invention under training with different known numbers of categories. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0039] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those generally understood by those skilled in the art of the technical field of the present invention. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. It is understood that the terms "first", "second", etc. used in the present invention can be used to describe various elements in this article, but these elements are not limited by these terms. These terms are only used to distinguish the first element from another element. For example, without departing from the scope of the present application, the first element can be referred to as the second element, and similarly, the second element can be the first element. Both the first element and the second element are elements, but they are not the same element.

[0040] In the field of recognition, there have been some works on collaborative recognition based on multi-model fusion. However, due to multiple factors such as model structure limitations and incomplete training sets, a single model often cannot achieve stable performance in real tasks. Multi-model fusion is to fuse multiple models together through some mechanism to obtain more powerful and stable performance than a single model. In the field of recognition, there have been some works on collaborative recognition based on multi-model fusion. Multi-model collaborative recognition can be roughly divided into several categories according to the level of fusion: (1) sensor level fusion, that is, data level fusion; (2) feature level fusion; (3) decision level fusion. Decision level fusion is a relatively high-level fusion level with low data transmission requirements and strong anti-interference ability. It can achieve stable results in reality. Therefore, the research on decision level fusion is relatively the most.

[0041] There are several typical decision fusion methods: (1) Weighted fusion methods, which essentially add a weight to the decision of each model for fusion. For example, through arithmetic mean fusion, set average fusion, and voting methods, these methods are relatively simple, easy to implement, and highly practical. (2) Methods based on ensemble learning ideas, such as bagging-based, boosting-based, and stacking-based fusion methods. This type of method emphasizes the formation of a strong classifier by combining various weak classifiers, and generally weak classifiers are homogeneous. However, it has certain constraints on the selection of training sets, training methods, and data interaction. Weak classifiers are difficult to interact with each other and are more suitable for multiple weak classifiers to be local, and are not well suited to distributed conditions. (3) Methods based on neural network ideas. This type of method realizes multi-model decision fusion by introducing neural networks. This type of method relies on training, and the generalization ability of the model is not strong.

[0042] Since the weighted fusion method is relatively simple, easy to implement, and has obvious effects in practical applications, it is widely used. The core idea of ​​the weighted fusion method is to combine the prediction results of different models according to certain weights to obtain a more accurate and robust overall prediction. The following is the general formula of the weighted fusion method:

[0043] score=α1model1+α2model2+…+α n model n (1)

[0044] Among them, score is the final fusion score, α i is the weight of the i-th model, model i is the prediction score of the ith model.

[0045] However, it should be noted that: overall, the current decision fusion method still has serious defects. It mainly manifests in two aspects: First, the decision fusion method of each model is rigid, and the weights are usually model-sensitive, but this weight ignores the fact that the recognition capabilities of different models under different tasks and different data are inconsistent. Second, it is not adaptable enough to distributed scenarios; the current decision fusion represented by ensemble learning does not fully consider the characteristics of distributed scenarios, and there is a lack of effective interaction mechanism between the models. However, distributed scenarios generally have no central node or no fixed central node. Under this limitation, the current method cannot be effectively applied.

[0046] In view of the problems existing in the prior art, in this embodiment, a multi-agent trusted collaborative target recognition method based on uncertainty quantification in an open environment is proposed. Figure 1-2 As shown, the method includes:

[0047] S1. Create a distributed communication network for multiple agents in an open environment, and each agent is independently loaded with its own corresponding target recognition model;

[0048] S2. Based on the uncertainty quantification mechanism, target recognition is performed on the task to be recognized through each target recognition model, and the corresponding target recognition results are obtained, wherein the target recognition results at least include: target recognition prediction data and uncertainty data, wherein the uncertainty data is used to characterize the uncertainty degree of the model for the target recognition prediction data, i.e., the reliability;

[0049] S3. Based on the multi-agent consistency mechanism, each agent is controlled to perform a first global information exchange through the communication path provided by the distributed communication network, so as to exclude exceptional data (also referred to as outliers or outliers) in the task to be identified through the first global information exchange, wherein the information participating in the first global interaction is the uncertainty data;

[0050] S4. Based on the multi-agent consistency mechanism, each agent is controlled to perform a second global information exchange through the communication path provided by the distributed communication network, so as to obtain the final recognition prediction result through the second global information exchange, wherein the information participating in the second global interaction is the target recognition prediction data.

[0051] Based on the above design scheme, it can be seen that the present application proposes a multi-agent trusted collaborative target recognition method based on uncertainty quantification (abbreviated as UQ-MCTR). The framework supports different agents to collaboratively achieve decision consensus under distributed conditions and complete recognition tasks; specifically, firstly, it performs recognition and analysis under distributed conditions through a multi-agent collaborative decision fusion framework; secondly, by introducing an uncertainty quantification mechanism (an uncertainty assessment is performed for each decision of each recognition model, and the uncertainty is the degree of sensitivity to the specific recognition task, that is, the uncertainty value reflects the "confidence" of the specific model in the specific decision), the UQ-MCTR method has the ability to distinguish out-of-distribution samples in open scenes (traditional collaborative target recognition methods are difficult to handle out-of-distribution (OOD: (Out-Of-Distribution)) targets); finally, by introducing a multi-agent consistency mechanism, a global information exchange and transmission mechanism is carried out between models, and the decisions of each agent are integrated, which improves the support for distributed multi-agents in an open environment to achieve decision consistency for target recognition tasks through this framework, and ensures the credibility of collaborative recognition decisions. Experiments have shown that on the CIFAR-10 dataset, the UQ-MCTR method is superior to the average voting method and the OOD single model method in terms of accuracy, AUROC, and AUPR, indicating that the UQ-MCTR method has obvious advantages in dealing with open environment problems and improving the generalization ability of the model.

[0052] In some specific embodiments, the specific steps of creating a distributed communication network for multiple agents in an open environment include:

[0053] Step S11: Assume that there are N agents participating in the same recognition task, that is, jointly identifying the target X to be recognized, where each agent is independently loaded with its corresponding target recognition model M. i , then the N target recognition models are uniformly expressed as M = {M1,M2,…,M i ,…,M N It should be noted that this application does not specify the target recognition model M of each intelligent agent. i The target recognition algorithm used is specifically limited, and the algorithms conventionally used in the field can be applied to this application, such as VGG, ResNet, YOLO, SSD and other target recognition network structures; the intelligent agent can be an unmanned boat;

[0054] Step S12: Create a distributed communication network in an open environment to provide distributed communication paths for each agent to interact with data. Specifically, in the scenario of multi-model fusion, each model is distributed into a directed graph, which can efficiently manage and coordinate the interaction, data flow and result integration process between different models. Therefore, this step adopts the method of obtaining the directed graph corresponding to each agent and obtaining the adjacency matrix corresponding to the directed graph G to form a distributed communication network. In a distributed system, each node in the directed graph not only describes the logical data flow, but also implies the allocation of computing resources and communication requirements, which helps to optimize resource utilization and reduce delays. Each node (corresponding to the target recognition model or processing step here) needs to follow a common communication protocol to exchange information. Each agent (target recognition model M i ) are all regarded as a node in the graph, and the process of transferring data or features from one model to another through directed edges. It can be direct feature transfer or model output (such as probability distribution, feature vector) as the input of the next model; specifically: considering that under distributed conditions, the data of X detected by each recognition model may be inconsistent due to differences in position, angle, platform, etc. (that is, even if the same target recognition algorithm is used, there will be certain differences). i The input data is represented as x i , the superscript i represents the i-th model. Assume that the target recognition model M i The corresponding output function is expressed as Its output can be expressed as It is assumed that each target recognition model is distributed to form a directed graph G, G = (V, E), where V = {v1, ..., v N} is a non-empty finite set of nodes, E is a set of edges; for a directed edge (v i ,v j ), side v i is called the parent node, v j is called a child node, and v i Yes j Neighbor; when (v i ,v j )∈E and (v j ,v i )∈E, it means G is an undirected graph; the node To Node The path is a sequence of edges in order, expressed as If there is a root node in a directed graph that has no parent and has directed paths to all other nodes in the graph, then the directed graph will contain a directed spanning tree where the node has no parent and therefore has directed paths to all other nodes in the graph.

[0055] At the same time, the adjacency matrix of the directed graph G is represented as A = [a ij ]∈R N×N For element a in adjacency matrix A ij , if (v j ,v i )∈E, then a ii =0,a ij =1, otherwise a ij = 0; Laplace matrix L = [l ij ]∈R N×N Defined as L ii =∑ j≠ i a ij ,L ij =-a ij , i≠j; let D=[d ij ]∈R N×N is a row random matrix associated with G, and when (v j ,v i )∈E then d ii >0,d ij >0, otherwise d ij =0; the 0 corresponding to the eigenvector 1 is a characteristic root of L, and all non-zero characteristic roots have positive real parts.

[0056] In some specific embodiments, the overall design concept of S2 is:

[0057] Object recognition plays a vital role in modern technology, especially in the fields of national defense security, intelligent surveillance, and autonomous driving. However, current object recognition technology faces several challenges, including handling unknown objects in open environments, improving the generalization ability of models, and enhancing the credibility of decisions.

[0058] In order to solve these problems, the present application proposes a multi-agent collaborative target recognition method based on uncertainty quantification, namely, an uncertainty quantification mechanism. Each agent initially performs preliminary recognition of the target through its corresponding target recognition model; in practical applications, in a distributed communication network, although the target recognition model of each agent can recognize the same recognition category, the differences in the training sets involved in the training will cause certain differences in the recognition accuracy of each category. For example, if the training set of the local node A located at the distributed position 1 contains more shore-based target samples and fewer sea mobile targets, the trained target recognition model 1 has a higher recognition accuracy for shore-based targets than for sea mobile targets. If the training set of the sea node B located at the distributed position 2 contains more sea mobile targets and fewer shore-based target samples, the trained target recognition model 2 has a lower recognition accuracy for shore-based targets than for sea mobile targets. Therefore, it is necessary to perform multi-model fusion on the recognition capabilities of multiple recognition models to obtain reliable results. At the same time, the applicant finds that uncertainty quantification has a very important application value in many fields, especially in fields with high decision-making risks. However, the current standard deep neural network cannot quantitatively evaluate uncertainty. For example, in the typical "cat-dog" recognition and classification task in the field of machine vision, when using blurred cat or dog images and non-cat and non-dog images (such as "pig") that the model has never seen, the target recognition model will still give a classification decision for the image - the probability value of a cat or dog. In these cases, the decision results given by the model are not credible, that is, there is a certain "uncertainty"; therefore, it is necessary to obtain the confidence level of each agent for the same target in different categories - the confidence level through the uncertainty quantification mechanism and exchange data through a distributed communication network (multi-agent consensus framework), that is, allow agents to exchange information through a well-defined communication topology (the distributed network depicts the communication path between agents), and the exchanged data is the probability distribution of target recognition, that is, the confidence level of each agent for the same target in different categories - the confidence level. In this way, the system aggregates the observation data of multiple agents while maintaining communication efficiency and significantly improving the accuracy of target recognition.

[0059] Based on the above design ideas, the specific steps of S2 include:

[0060] Through the uncertainty quantification mechanism, namely the uncertainty quantification method MC-dropout, the posterior distribution form of the network parameters of the target recognition model is changed (thereby giving the uncertainty of the output result), and the same input data is tested multiple times to obtain the target recognition results of each target recognition model for the task to be recognized, and the target recognition results at least include: target recognition prediction data and uncertainty data, the uncertainty data is used to characterize the uncertainty degree of the model for the target recognition prediction data, that is, the reliability; the target recognition prediction data is the probability distribution data corresponding to the target recognition model for different recognition categories;

[0061] Specifically, assume that each recognition model M i , the recognition model M i The recognition ability is to be able to identify q-type targets, and its corresponding input data x i , identify the network parameters as Then for the input data x i , test H times, each test corresponds to the output data is a (1x q) vector, and the corresponding sign function is h∈(1, 2, ..., H), the final target recognition result is the target recognition prediction data and uncertainty data, where the target recognition prediction data is the recognition model M i The predicted class probability distribution E(y) corresponding to the target identified in the current task is a certain category i ), the specific probability distribution calculation process is not limited, and its expression can be represented by the following formula

[0062]

[0063] The uncertainty data Var(y i ) is the prediction variance, which means the identification model M i The uncertainty measure for this inference can be expressed as follows:

[0064]

[0065] Among them, τ is the additive noise term, which represents the inherent uncertainty of the model, and its expression is τ = l 2 p / 2Nλ, l is the prior length scale, p is the probability that each neural unit is not abandoned during the Dropout process, N is the number of input samples, λ is the regularization weight attenuation, I D is the unit matrix to ensure the consistency of the front and back dimensions; Expresses the random vector y i The estimate of E(y i )T E(y i ) is a random vector y i The expected value E(y i ) (variance estimate), y i For output data A random vector in

[0066] Through the above uncertainty quantification mechanism, the probability distribution corresponding to each recognition category and the corresponding result of uncertainty can be obtained, which is expressed as c i ={E(y i ),Var(y i )}.

[0067] In some specific embodiments, in order to achieve overall decision consistency from the cluster level in this application, it is necessary to communicate between multiple agents to reach a consistent result. Therefore, it is necessary to extract the probability distribution and probability mean obtained by a single model and exchange data to achieve multi-model fusion, specifically referring to each target recognition model M i After obtaining their respective preliminary prediction results c i Finally, in order to achieve overall decision consistency at the cluster level, two global data exchanges are required, and the specific implementation process is completed by S3-S4.

[0068] The implementation process of S3 includes: based on the multi-agent consistency mechanism, through the communication path provided by the distributed communication network, controlling each agent to perform the first global information exchange, so as to exclude the exception data in the task to be identified through the first global information exchange, wherein the information participating in the first global interaction is the uncertainty data; the control process in which the multi-agent consistency mechanism participates in this step is:

[0069] S31, setting global information exchange constraints and initializing;

[0070] The setting of global information exchange constraints specifically refers to setting the communication starting node t0 and the number of times n that global communication is completed between nodes, thereby controlling the time of information exchange between multiple agents;

[0071] Initialization specifically refers to setting and initializing an array A for storing the state solutions of each multi-agent at each communication time point;

[0072] S32, forming a communication feature vector corresponding to each intelligent agent, and controlling each intelligent agent to perform the first global information exchange through the communication path provided by the distributed communication network, that is, sending the local communication feature vector to the neighboring node and receiving the communication feature vector sent by the neighboring node, thereby obtaining the uncertainty fusion result after the first global information exchange In the first global information exchange, each agent exchanges information with its corresponding neighboring nodes n times, and in each information exchange process, a consensus algorithm is used to calculate the communication result after each exchange of each agent, i.e., the uncertainty;

[0073] The communication feature vector is used to characterize the first initial state corresponding to each agent, and the element of the feature vector is the target recognition model M i The preliminary prediction results obtained i The uncertainty in ; that is, the uncertainty is set as a parameter that characterizes the initial state of the agent;

[0074] After forming the communication feature vector corresponding to each intelligent agent, each intelligent agent performs n information exchanges or information exchanges with its corresponding neighboring nodes through the communication path (communication topology diagram) provided by the distributed communication network, so that in each exchange, the local communication feature vector is sent to the neighboring node and the communication feature vector sent by the neighboring node is received. Finally, the uncertainty fusion result corresponding to each recognition category after n information exchanges is obtained through the continuous time consistency algorithm.

[0075] The formula corresponding to the continuous time consistency algorithm is:

[0076]

[0077] Among them, x i (t), x j (t) is the information to be exchanged between agent i and its neighboring node, agent j, i.e., uncertainty, l ij is the i-th row and j-th column element of the Laplace matrix L; if the communication topology graph contains a directed spanning tree, then we know that t→∞, where v j is the n×1 non-negative left eigenvector of the Laplace matrix L corresponding to the 0 eigenvalue;

[0078] Since for each data, each recognition model M i Each of them will undergo an MC-dropout process to estimate the uncertainty of the data. If we assume that the corresponding uncertainty is Var i (y0), then each recognition model M i The estimated value of the data is used as the initial value of the consistency algorithm and is brought into the corresponding x i (t), after n rounds of iterations based on formula (4), the final uncertainty fusion result can be obtained

[0079] S33, based on the given threshold value, the uncertainty fusion result Screening is performed to exclude the exceptional data, i.e., OOD data, in the task to be identified; specifically, in order to correctly screen out the OOD data, the uncertainty fusion results need to be screened according to the threshold. Therefore, a threshold discrimination mechanism is introduced for out-of-distribution detection, namely:

[0080]

[0081] Among them, θ is the decision threshold, which is used to judge the input data x i Whether a certain point x to be identified or a sample to be identified is outside the distribution, that is, whether the category corresponding to the sample exceeds the target recognition model M i The categories that can be identified are When it is not less than θ, it means that after the above consistency fusion analysis, it is unanimously believed that the point x to be identified is Out of Distribution (OOD) data, and the OOD data will be rejected for category recognition. i The above results show that the application can fuse the results through uncertainty. Further screen the test data, remove the test data with uncertainty greater than the decision threshold θ, and only leave the test data with uncertainty less than the decision threshold θ and the corresponding category label. Among them, the decision threshold θ determines the accuracy of the retained test data, that is, the proportion of ID data in the retained test data. If the decision threshold is too high, the proportion of OOD data in the retained test data will increase, resulting in a decrease in accuracy; if the decision threshold is too low, a large amount of ID data will be removed, resulting in deviations in the prediction results. Therefore, it is necessary to select a suitable decision threshold θ. In this application, it is recommended that the decision threshold θ be selected as 1.

[0082] S4. Based on the multi-agent consistency mechanism, each agent is controlled to perform a second global information exchange through the communication path provided by the distributed communication network, so as to obtain the final recognition prediction result through the second global information exchange, wherein the information participating in the second global interaction is the target recognition prediction data. The control process involved in the multi-agent consistency mechanism in this step is:

[0083] S41, forming a communication feature matrix corresponding to each intelligent agent, and controlling each intelligent agent to perform a second global information exchange through the communication path provided by the distributed communication network, that is, sending the local communication feature matrix to the neighboring node and receiving the communication feature matrix sent by the neighboring node, thereby obtaining the final prediction fusion result after the second global information exchange, that is, the fusion result of the prediction class probability distribution In the second global information exchange, each agent exchanges information with its corresponding neighboring nodes n times, and in each information exchange process, a consensus algorithm is used to calculate the communication results of each agent after each exchange, that is, the probability distribution;

[0084] The AC feature matrix is ​​used to characterize the second initial state corresponding to each agent. In the elements of the AC feature matrix, each row corresponds to the test set-screened input data x i The classification result of each image sample in is the target recognition model M i The respective prediction results c i The probability values ​​in each column correspond to the category specified by the data set; then after forming the communication feature matrix corresponding to each intelligent agent, each intelligent agent performs n information exchanges or information exchanges with its corresponding neighboring nodes through the communication path (communication topology diagram) provided by the distributed communication network, so that in each exchange, the local communication feature matrix is ​​sent to the neighboring node and the communication feature matrix sent by the neighboring node is received, and finally, the prediction fusion result corresponding to each recognition category after n information exchanges is obtained by the continuous time consistency algorithm;

[0085] The formula corresponding to the continuous time consistency algorithm is:

[0086]

[0087] Among them, x i (t), x j (t) is the information to be exchanged between agent i and its neighboring node, agent j, i.e., the probability distribution, l ij is the i-th row and j-th column element of the Laplace matrix L; if the communication topology graph contains a directed spanning tree, then we know that t→∞, where v j is the n×1 non-negative left eigenvector of the Laplace matrix L corresponding to the 0 eigenvalue;

[0088] Since for each data, each recognition model M i A MC-dropout process will be performed to estimate the probability distribution of the data. If the corresponding probability distribution is assumed to be E i (y0), then each recognition model M i The estimated value of the data is used as the initial value of the consistency algorithm and is brought into the corresponding x i (t), after n rounds of iterations based on formula (4), the final uncertainty fusion result can be obtained.

[0089] S42, the prediction fusion results corresponding to each identification category after the obtained n information exchanges are input into the final classification layer, and the prediction labels of the data are obtained to complete the information exchange between multiple models, and then the final c i , where the final classification layer is the softmax layer, which directly outputs the category with the highest probability in the predicted fusion result.

[0090] This application also verifies the above design scheme through actual cases, and the corresponding experimental scheme is:

[0091] A. Experimental Setup

[0092] a. Introduction to the test set used - CIFAR-10 dataset

[0093] like Figure 3 This experiment uses the classic image classification data CIFAR-10 dataset, which was released by the Canadian Institute for Advanced Research in 2009 to evaluate the performance of deep generative models. Specifically, CIFAR-10 is a subset of the original Tiny Images dataset, containing 60,000 32×32 pixel color images. The dataset is divided into 10 categories, with 6,000 images in each category. Among them, 50,000 images are used for training and 10,000 images are used for testing.

[0094] b. UQ-MCTR experimental results and comparative experimental results

[0095] In order to determine the advantages of the solution proposed in this application in an open environment, this experiment uses the average voting method as a comparison method for ensemble learning. At the same time, a comparative analysis is performed with a single model method with OOD detection (three deep learning convolutional neural network models ResNet50, VGG16 and ResNet18) to prove the effectiveness of the solution proposed in this application. Among them, ResNet50 and ResNet18 are residual networks. By using jump connections, the degradation problem of deep networks is solved, and the performance and efficiency of the network are improved; and VGG16 is a visual geometry group network proposed by Simonyan and Zisserman. By using multiple 3×3 convolution kernels, a balance between depth and width is achieved, and the expression and generalization capabilities of the network are improved. As a classic ensemble learning algorithm, the average voting method improves the overall performance by combining the predictions of multiple models, reduces the variance of the model, and thus reduces the risk of overfitting.

[0096] At the same time, this experiment uses accuracy, AUPR, and AUROC as evaluation indicators of model performance. Accuracy is the ratio of the number of samples correctly predicted by the model to the total number of samples. Considering the uncertainty of the open environment, the accuracy in the open environment is ACC. ORedefined as:

[0097]

[0098] Among them, the symbol correspondence table is shown in Table 5.1.

[0099] Table 5.1 Symbol correspondence table

[0100] TP The number of correctly classified positive samples. TN The number of negative samples that are correctly classified. FP The number of misclassified positive samples. FN The number of misclassified negative samples. TU Samples correctly excluded based on uncertainty. FU Samples incorrectly excluded based on uncertainty.

[0101] AUROC can reflect the model's ability to distinguish between positive and negative samples under different thresholds, that is, the balance between the true positive rate (TPR) and the false positive rate (FPR). The closer the AUROC value is to 1, the better the classification effect of the model. The calculation formula of AUROC can be expressed as:

[0102] AUROC=P(score positive >score negative )

[0103] Among them, score positive and score negative Represent the scores of positive samples and negative samples respectively.

[0104] AUPR is the area under the precision-recall curve and is an indicator for evaluating the performance of a binary classification model. AUPR calculates the relationship between precision and recall under different decision thresholds and is generally used to evaluate the model's ability to predict positive samples. If a model has a high AUPR and a low accuracy, it means that the model has a strong ability to recognize minority classes, but a weak ability to recognize majority classes. If a model has a low AUPR and a high accuracy, it means that the model has a strong ability to recognize majority classes, but a weak ability to recognize minority classes.

[0105] c. Experimental parameter settings

[0106] In this experiment, each of the above models was implemented using the Tensorflow and Keras frameworks, and trained and tested on the cifar-10 dataset. The specific settings are: use 100 epochs to train the model, the batch size is uniformly set to 256, and the OOD decision threshold is set to 1.0. To ensure a consistent operating environment, the Python version is 3.8, the computer graphics card is 4060, and CUDA is used for acceleration during the training and testing process.

[0107] B. Experimental Results

[0108] a. Experimental analysis

[0109] In order to simulate the scenario of open set recognition, this application selected 7 categories as known categories and the remaining 3 categories as unknown categories, conducted experiments and obtained Table 5.2. As shown in the table, the UQ-MCTR method improves the accuracy by 29.7% compared with the average voting method, and improves by 0.156 and 0.383 in AUROC and AUPR respectively. This shows that the performance of this application in an open environment is better than the average voting method, and the unknown categories can be correctly excluded. For a single model, after adding uncertainty, the performance of the three models (Resnet18, Resnet50 and VGG16) has been significantly improved, and each model can make certain judgments on unknown categories in an open environment. However, compared with the single model with the addition of uncertainty, this application has a significant performance improvement in terms of accuracy, AUROC and AUPR. This shows that this application can effectively utilize the information of multiple agents, improve the accuracy of classification, and effectively improve the overall performance of the model.

[0110] Table 5.2UQ-MCTR and comparative experimental results

[0111]

[0112] At the same time, the application also studies the impact of different threshold values ​​on the application, and the application conducts further experiments on the model under different OOD decision thresholds. Figure 4 It can be seen that the overall trend of accuracy is an increase and then decrease, with a peak value of 1.0. The overall trend of AUPR and AUROC is a downward trend. Experiments show that under different uncertainties, the accuracy is affected by the number of uncertain excluded samples. At low uncertainty thresholds, a large number of known class samples are excluded, resulting in an increase in FU value, thereby reducing the accuracy. At high uncertainty thresholds, the number of samples correctly excluded by uncertainty gradually decreases, resulting in a decrease in TU. Therefore, the overall trend is an increase and then a decrease. This shows that under different threshold values, its performance will be affected to a certain extent. It is necessary to comprehensively consider the trade-offs in performance to obtain an ideal recognition model.

[0113] In addition, this application also studies the impact of different OOD settings. Since the OOD setting (known category) may affect the model performance during model training, this application conducts experiments on models under different OOD settings. In the experiments with different OOD settings, the average voting method is compared and the following results are obtained: Figure 5The experimental results are shown in Figure 2. In the experiments with known categories, as the number of known categories increases, the accuracy, AUPR, and AUROC all show significant improvements. For example, when the number of labels is 10, the accuracy of known label training reaches 0.964, while the accuracy of the average voting method is 0.831. This difference is also obvious in AUPR and AUROC. This shows that the more known categories there are during training, the more accurate and reliable results it can provide.

[0114] In summary, the performance of this application in an open environment is better than the average voting method, indicating that this application can accurately distinguish unknown categories in an open environment. Compared with the single model OOD method, this application also has significant improvements in performance. This shows that this application has significant advantages in dealing with open environment uncertainty and improving model generalization capabilities.

[0115] Based on the same inventive concept, the present invention also proposes a computer-readable storage medium, comprising computer instructions, and when the computer instructions are executed on a computer, the computer executes the method described.

[0116] In summary, the implementation of the embodiments of the present invention will have the following beneficial effects:

[0117] The multi-agent trusted collaborative target recognition method based on uncertainty quantification proposed in this application has shown significant innovation and superiority in the field of machine learning image recognition. A multi-agent distributed decision fusion framework is constructed through a distributed communication network, which achieves decision consistency for target recognition tasks in an open environment and improves the credibility of recognition decisions. In particular, in a multi-agent system, the proposed consistency mechanism effectively achieves decision consensus and ensures the robustness and flexibility of the system. Experimental results show that the performance of the UQ-MCTR method in an open environment is better than the average voting method, and it can accurately distinguish unknown categories. Compared with the single-model OOD method, the multi-agent OOD method has significant improvements in accuracy, AUROC and AUPR, proving its significant advantages in dealing with open environment uncertainty and improving model generalization ability. It shows that the UQ-MCTR method has significant advantages in dealing with open environment uncertainty and improving model generalization ability. Future work will focus on further improving the robustness of the model, as well as exploring more multi-agent consistency techniques and uncertainty quantification techniques to enhance the applicability and reliability of the model in a wider range of open environments.

[0118] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A multi-agent credible collaborative target recognition method based on uncertainty quantification in an open environment, characterized by: include: S1. Create a distributed communication network for multiple agents in an open environment, and each agent is independently loaded with its own corresponding target recognition model; S2. Based on the uncertainty quantification mechanism, target recognition is performed on the task to be recognized through each target recognition model, and the corresponding target recognition results are obtained, wherein the target recognition results at least include: target recognition prediction data and uncertainty data, wherein the uncertainty data is used to characterize the uncertainty degree of the model for the target recognition prediction data, i.e., the reliability; S3. Based on the multi-agent consistency mechanism, each agent is controlled to perform a first global information exchange through the communication path provided by the distributed communication network, so as to exclude the exception data in the task to be identified through the first global information exchange, wherein the information participating in the first global interaction is the uncertainty data; S4. Based on the multi-agent consistency mechanism, each agent is controlled to perform a second global information exchange through the communication path provided by the distributed communication network, so as to obtain the final recognition prediction result through the second global information exchange, wherein the information participating in the second global interaction is the target recognition prediction data.

2. The multi-agent trusted collaborative target recognition method based on uncertainty quantification in an open environment according to claim 1 is characterized in that: In S2, the posterior distribution form of the network parameters of the target recognition model is changed through the uncertainty quantification mechanism, namely the uncertainty quantification method MC-dropout, and multiple tests are performed on the same input data to obtain the target recognition results of each target recognition model for the task to be recognized, and the target recognition results at least include: target recognition prediction data and uncertainty data, and the uncertainty data is used to characterize the uncertainty degree of the model for the target recognition prediction data, that is, the reliability; the target recognition prediction data is the probability distribution data corresponding to the target recognition model for different recognition categories.

3. The multi-agent trusted collaborative target recognition method based on uncertainty quantification in an open environment according to claim 2 is characterized in that: S2 also gives the following assumptions, which include: Assume that each recognition model M i , the recognition model M i The recognition ability is to be able to identify q-type targets, and its corresponding input data x i , identify the network parameters as Then for the input data x i , test H times, each test corresponds to the output data is a (1, q) vector, and the corresponding sign function is h∈{1, 2, …, H}, then the final target recognition result is the target recognition prediction data and uncertainty data. The target recognition prediction data is the i The predicted class probability distribution E(y i ); uncertainty data Var(y i ) is the prediction variance, which means the identification model M i For the uncertainty measurement of target recognition prediction data, the probability distribution corresponding to each recognition category and the result corresponding to the uncertainty degree can be obtained through the above uncertainty quantification mechanism, which is expressed as c = {E(y i ),Var(y i )}.

4. The multi-agent trusted collaborative target recognition method based on uncertainty quantification in an open environment according to claim 1 is characterized in that: The specific steps of S3 include: S31, setting global information exchange constraints and initializing; The setting of global information exchange constraints specifically refers to setting the communication starting node t0 and the number of times n that global communication is completed between nodes, thereby controlling the time of information exchange between multiple agents; Initialization specifically refers to setting and initializing an array A for storing the state solutions of each multi-agent at each communication time point; S32, forming a communication feature vector corresponding to each intelligent agent, and controlling each intelligent agent to perform the first global information exchange through the communication path provided by the distributed communication network, that is, sending the local communication feature vector to the neighboring node and receiving the communication feature vector sent by the neighboring node, thereby obtaining the uncertainty fusion result after the first global information exchange In the first global information exchange, each agent exchanges information with its corresponding neighboring nodes n times, and in each information exchange process, a consensus algorithm is used to calculate the communication results of each agent after each exchange, i.e., the uncertainty; The communication feature vector is used to characterize the first initial state corresponding to each agent, and the element of the feature vector is the target recognition model M i The preliminary prediction results obtained i The uncertainty in ; that is, the uncertainty is set as a parameter that characterizes the initial state of the agent; After forming the communication feature vector corresponding to each intelligent agent, each intelligent agent performs n information exchanges or information exchanges with its corresponding neighboring nodes through the communication path provided by the distributed communication network, so that in each exchange, the local communication feature vector is sent to the neighboring node and the communication feature vector sent by the neighboring node is received, and finally the uncertainty fusion result corresponding to each recognition category after n information exchanges is obtained. S33, based on the given threshold value, the uncertainty fusion result Screening is performed to exclude exceptional data, namely OOD data, in the tasks to be identified.

5. The multi-agent trusted collaborative target recognition method based on uncertainty quantification in an open environment according to claim 4 is characterized in that: The specific steps of S4 include: S41, forming a communication feature matrix corresponding to each intelligent agent, and controlling each intelligent agent to perform a second global information exchange through the communication path provided by the distributed communication network, that is, sending the local communication feature matrix to the neighboring node and receiving the communication feature matrix sent by the neighboring node, and then obtaining the final prediction fusion result after the second global information exchange, that is, the fusion result of the prediction class probability distribution In the second global information exchange, each agent exchanges information with its corresponding neighbor nodes n times, and in each information exchange process, a consensus algorithm is used to calculate the communication results of each agent after each exchange, that is, the probability distribution value; The AC feature matrix is ​​used to characterize the second initial state corresponding to each agent. In the elements of the AC feature matrix, each row corresponds to the test set, i.e., the screened input data x i The classification result of each image sample in is the target recognition model M i The respective prediction results c i The probability values ​​in , each column corresponds to the category specified by the test set; then after forming the communication feature matrix corresponding to each intelligent agent, each intelligent agent exchanges information with its corresponding neighboring node n times through the communication path provided by the distributed communication network, so that in each exchange, the local communication feature matrix is ​​sent to the neighboring node and the communication feature matrix sent by the neighboring node is received, and finally the prediction fusion result corresponding to each recognition category after n information exchanges is obtained; S42, the prediction fusion results corresponding to each recognition category after the obtained n information exchanges are input into the final classification layer, and the prediction labels of the data are obtained to complete the information exchange between multiple models, and then the final c i .

6. The multi-agent trusted collaborative target recognition method based on uncertainty quantification in an open environment according to claim 1, characterized in that: In S1, a distributed communication network is formed by obtaining a directed graph corresponding to each intelligent agent and obtaining an adjacency matrix corresponding to the directed graph G.

Images