Object detection model causal decoupling adaptive evaluation method and device

Abstract

The present application relates to a kind of target detection model cause and effect decoupling adaptive evaluation method and device, belong to target detection and artificial intelligence field.The evaluation device of the present application includes: cause and effect migration evaluation reasoning module, architecture-strategy coupling degree differentiable measure module and hyperbolic space quality embedding module;Cause and effect migration evaluation reasoning module is used to build the cause and effect dependent relationship of scene and strategy, generates counterfactual sample, and filters out the robust strategy of anti-environmental interference;Architecture-strategy coupling degree differentiable measure module is used to establish the depth coupling relationship of evaluation strategy and model architecture, and realizes the automatic search of strategy by differentiable optimization;Hyperbolic space quality embedding module is used to solve the metric distortion problem of hierarchical strategy using non-euclidean geometry space, and the confidence of quantitative evaluation result.The present application can reduce the cost of artificial parameter adjustment, improve the reliability of evaluation result in complex environment.

Description

Technical Field

[0001] This invention belongs to the fields of object detection and artificial intelligence, and specifically relates to an adaptive evaluation method and apparatus for causal decoupling of object detection models. Background Technology

[0002] With the rapid development of deep learning and artificial intelligence technologies, intelligent decision-making systems, represented by object detection, have been widely applied in important fields such as perimeter security, autonomous driving, and military intelligence in recent years. Therefore, the accuracy and robustness of object detection model performance evaluation are becoming increasingly important. Currently, object detection evaluation tasks in open and dynamic environments face severe challenges such as large environmental distribution shifts, difficulty in adjusting evaluation strategy parameters, and difficulty in quantifying confidence in long-tail / extreme scenarios. Therefore, it is necessary to consider using cutting-edge technologies such as causal inference and geometric representation learning to construct an intelligent evaluation system that can adapt to environmental changes and automatically search for optimal parameters, tailing intelligent decision-making models to these systems. Traditional evaluation methods often rely on static test datasets and fixed evaluation metrics, which not only struggle to cope with the interference of complex factors such as lighting, weather, and occlusion in the real world, but are also prone to learning spurious correlations due to dataset bias. Furthermore, evaluation strategies typically include a large number of hyperparameters such as non-maximum suppression (NMS) thresholds and anchor box sizes; traditional manual tuning methods are inefficient and cannot guarantee optimality. Therefore, using adaptive learning technology to construct intelligent evaluation methods can not only effectively eliminate the interference of environmental confounding factors on evaluation results, but also achieve the co-evolution of evaluation strategies and model architecture, providing a strong security guarantee for the deployment of intelligent decision-making systems.

[0003] The evaluation strategy for object detection models refers to a comprehensive algorithmic logic consisting of a series of post-processing parameters, evaluation index weights, and confidence thresholds, under specific application scenario constraints. It describes how to transform the model's raw output into the final decision result and measure its quality. Intelligent evaluation techniques for object detection assess the model's discriminability, stability, and accuracy in different distributed scenarios by constructing causal dependencies and searching for optimal parameter combinations. The evaluation strategy determines the model's performance in actual deployment and serves as a bridge connecting model training and practical application. Currently, common methods in this field mainly include: static evaluation based on statistical indicators, model optimization based on neural architecture search, and stress testing based on large-scale synthetic data. With the development of automated machine learning technology, methods using meta-learning to assist in adjusting evaluation parameters have also emerged, improving evaluation efficiency to some extent. However, because traditional methods are mostly based on Euclidean space for measurement and lack consideration of data causal mechanisms, they suffer from severe distance distortion when dealing with complex policy spaces with hierarchical structures, and cannot distinguish whether the model has truly learned object features or merely fitted background noise. Meanwhile, the above methods still have problems such as the evaluation criteria not being able to be dynamically adjusted with the architecture, the lack of mathematical measurement of uncertainty, and poor cross-scenario transferability. There are few intelligent evaluation methods for target detection models that combine causal inference and hyperbolic geometric embedding.

[0004] Evaluation techniques for intelligent decision-making models, especially object detection models, are a core element in ensuring the safe deployment of artificial intelligence systems. In this field, how to construct an evaluation system that can adapt to open and dynamic environments and accurately reflect the true capabilities of models has always been a research hotspot. Traditional evaluation methods mainly rely on static test sets and fixed statistical indicators. With the rise of automated machine learning and neural architecture search technologies, evaluation techniques are also beginning to develop towards automation and intelligence. Related patents and research technologies are as follows: Patent [CN112464960A: A Target Detection Method Based on Fast Neural Architecture Search] discloses a method for automatically customizing target detection network models using a search algorithm. This method first defines a search space consisting of a multi-layer feature fusion module and a detection head module, constructs a search algorithm, and applies it to the search space to search the original image, ultimately detecting the target's category and location information. This invention achieves model construction that balances efficiency and accuracy with limited computing resources through automated search. However, this invention only focuses on the search and optimization of the neural network architecture (Backbone / Neck / Head), and its evaluation stage still uses fixed post-processing strategies and discrete evaluation metrics, lacking means for adaptive search and differentiable optimization of the evaluation strategy itself, resulting in a disconnect between the searched architecture and the evaluation strategy.

[0005] Patent [CN114998707B: Attack Method and Apparatus for Evaluating the Robustness of an Object Detection Model] discloses a method for evaluating the robustness of an object detection model. This method calculates the importance of key pixels, iteratively attacks them, and evaluates the model's robustness based on the attack success rate. This approach, by dynamically adding and restoring perturbed key pixels, can more effectively uncover the model's vulnerabilities. However, the "robustness evaluation" employed in this invention is primarily based on pixel-level adversarial perturbations, focusing on defending against malicious attacks. It does not address the distribution shift problem commonly encountered in natural scenes from the perspective of "causal inference," and therefore cannot effectively evaluate the model's true generalization ability in cross-domain scenarios.

[0006] Patent [CN114419346B: A Method, Apparatus, Device, and Medium for Model Robustness Detection] discloses a general model robustness detection scheme. This method acquires adversarial example images, inputs them into the target model to be attacked, obtains the output results, and obtains robustness diagnostic results based on the output results and a preset similarity threshold. This invention can output objective robustness diagnostic data for the model. However, when measuring the distance or similarity between the model output and the evaluation criteria, this invention mainly relies on the traditional Euclidean space metric system. For evaluation strategy spaces with complex hierarchical structures, Euclidean metrics will produce severe geometric distortions, making it impossible to accurately quantify the quality distribution and uncertainty of the evaluation strategies.

[0007] Therefore, existing methods still have several limitations. First, the lack of causal mechanism-based evaluation methods makes it impossible to eliminate spurious correlations caused by environmental background, leading to cross-scenario evaluation failure. Second, model architecture search and evaluation strategy parameter optimization are disconnected, lacking a differentiable coupling metric mechanism to achieve end-to-end joint optimization. Finally, traditional Euclidean space metric methods cannot adapt to the hierarchical characteristics of evaluation strategies and lack precise geometric embedding and expression of evaluation quality uncertainty. Summary of the Invention

[0008] (a) Technical problems to be solved The technical problem this invention aims to solve is how to provide a causal decoupling adaptive evaluation method and apparatus for target detection models, in order to address the shortcomings of existing methods. These methods lack causal mechanism-based evaluation methods, cannot eliminate spurious correlation interference caused by environmental background, leading to cross-scene evaluation failure; model architecture search and evaluation strategy parameter optimization are in a disconnected state, lacking a differentiable coupling metric mechanism to achieve end-to-end joint optimization; and traditional Euclidean space metric methods cannot adapt to the hierarchical characteristics of evaluation strategies, lacking accurate geometric embedding and expression of evaluation quality uncertainty.

[0009] (II) Technical Solution To address the aforementioned technical problems, this invention proposes a causal decoupling adaptive evaluation device for a target detection model. The device includes: a causal transfer evaluation inference module, an architecture-policy coupling degree differentiable measurement module, and a hyperbolic space quality embedding module. The causal transfer evaluation and reasoning module is used to construct the causal dependency relationship between scenarios and policies, generate counterfactual samples, and screen out robust policies resistant to environmental interference; it includes three sub-modules: The scenario-policy dependency graph construction submodule includes a causal relationship discovery unit and a graph node definition unit. The causal relationship discovery unit uses the PC algorithm to scan the dataset metadata, mine the conditional independence between the scenario environment, model features and evaluation policy parameters, and construct a directed acyclic graph. The graph node definition unit defines environmental factors as confusion nodes and policy parameters as intervention nodes, clarifying the causal flow between variables. The counterfactual virtual scene generation submodule includes an intervention feature generation unit and a virtual sample synthesis unit. The intervention feature generation unit introduces a do-calculus intervention quantifier to sever the association between environmental nodes and target nodes in the feature space and apply perturbation to the environmental feature vector. The virtual sample synthesis unit generates unseen counterfactual virtual scene features based on the perturbed features to simulate OOD scenarios. The invariant risk minimization constraint submodule includes an environmental risk calculation unit and a robust policy selection unit. The environmental risk calculation unit calculates the evaluation loss of candidate policies in real scenarios and counterfactual virtual scenarios, respectively. The robust policy selection unit introduces environmental invariance constraints to eliminate policies that perform too differently under different environmental interventions and retains a subset of robust policies that focus on the essential characteristics of the target. The architecture-policy coupling degree differentiable measurement module is used to establish a deep coupling relationship between the evaluation policy and the model architecture, and to achieve automated policy search through differentiable optimization; it includes three sub-modules: The multidimensional orthogonal feature mapping submodule includes a coupled dimension calculation unit and a feature orthogonalization unit. The coupled dimension calculation unit decomposes the matching relationship between the model and the policy into four dimensions: performance response, information entropy, Jacobi sensitivity, and computational cost. The feature orthogonalization unit decouples the feature vectors of the above four dimensions to ensure that each dimension remains orthogonal and independent in the metric space. The scene-driven meta-weight generation submodule includes a scene feature extraction unit and an attention weight prediction unit. The scene feature extraction unit analyzes the distribution characteristics of the current data to be evaluated. The attention weight prediction unit uses a lightweight meta-network to dynamically predict the importance weight vectors of the above four dimensions based on the scene features. The policy parameter differentiable search submodule includes a parameter relaxation processing unit and a joint gradient update unit. The parameter relaxation processing unit uses the Gumbel-Softmax function to continuously relax the discrete evaluation policy parameters. The joint gradient update unit constructs a backpropagation path from the weighted evaluation index to the policy parameters, so that the evaluation policy can undergo joint gradient descent during the search process of the model architecture and converge to the optimal policy parameters. The hyperbolic space quality embedding module is used to address the metric distortion problem of hierarchical strategies using non-Euclidean geometry and to quantify the confidence of evaluation results; it includes three sub-modules: The hyperbolic manifold projection submodule includes a curvature adaptive learning unit and an exponential mapping unit. The curvature adaptive learning unit automatically finds the optimal manifold curvature that fits the current policy tree level complexity through gradient descent. The exponential mapping unit uses the tangent space as an intermediary to project the policy feature vector optimized in Euclidean space onto the hyperbolic manifold. On a spherical manifold, to preserve the inherent hierarchical semantic distance between policies; The policy quality decoupling encoding submodule includes a quality subspace partitioning unit and a manifold feature encoding unit. The quality subspace partitioning unit further decomposes the policy quality into four independent Riemann subspaces in hyperbolic space: stability, discriminability, robustness, and efficiency. The manifold feature encoding unit encodes the policy vector in each subspace to obtain a more refined quality representation. The Riemann uncertainty quantification submodule includes a Riemann Gaussian distribution estimation unit and a confidence level determination unit. The Riemann Gaussian distribution estimation unit calculates the probability distribution of the strategy on the hyperbolic manifold, outputs the mean of the distribution as the final quality score, and calculates the geodesic variance of the distribution. The confidence level determination unit uses the geodesic variance as an uncertainty measure. When the uncertainty exceeds a preset threshold, it automatically marks the evaluation result as "low confidence" and triggers a manual review alarm.

[0010] This invention also provides an adaptive evaluation method for causal decoupling of object detection models, the method comprising the following steps: S1. Preparation and Causal Discovery Phase: The system scans metadata to build a structural causal model, defines the dependencies between environment, goal, and policy, and constructs a scenario-policy causal graph; secondly, it uses the backdoor criterion to identify environmental confounding factors; finally, it generates a "counterfactual augmentation dataset" by performing counterfactual intervention in the feature space; and introduces an invariant risk minimization (IRM) mechanism. S2. Differentiable Policy Search Stage: First, the model and policy are mapped to four orthogonal Riemannian manifold spaces: performance, information, sensitivity, and efficiency. Second, the meta-weight network is used to dynamically generate attention weights for each dimension based on scene features. Third, the Gumbel-Softmax relaxation technique is introduced to make the discrete policy parameters continuous. Finally, end-to-end joint gradient descent is performed through bi-layer programming to search for the optimal policy configuration under the current architecture. S3, Hyperbolic Quality Embedding Stage: Projecting the policy vector in Euclidean space onto a hyperbolic space with constant negative curvature using an exponential mapping. Inside the spherical manifold; the manifold curvature is automatically learned through gradient descent to adapt to the hierarchical structure of the policy; finally, the policy is modeled as a Riemannian Gaussian distribution, and the quality score and uncertainty confidence are output. S4. Evaluation, Feedback and Decision-Making Stage: Transform mathematical evaluation results into engineering decisions; based on the uncertainty of Riemann variance determination, decide whether to adopt automatic evaluation results or trigger manual review; at the same time, feed back high-confidence difficult samples to the model training loop to complete the closed loop of "evaluation-feedback-optimization".

[0011] (III) Beneficial Effects This invention proposes a causal decoupling adaptive evaluation method and apparatus for target detection models, which has the following beneficial effects: 1. This invention proposes a causal transfer evaluation and reasoning technique based on scenario-policy dependency graph. It aims to eliminate spurious correlations between environmental background and target at the feature level by using a structural causal model and a counterfactual intervention mechanism. This effectively solves the technical problem of traditional evaluation methods failing to evaluate models in new cross-domain scenarios and having inaccurate generalization ability assessments due to biased training data distribution.

[0012] 2. This invention proposes a method for measuring the differentiable degree of architecture-policy coupling based on multi-view orthogonal feature fusion. It aims to establish a fully differential gradient link between evaluation policy parameters and model architecture by using meta-weighted networks and Gumbel-Softmax relaxation techniques, thereby achieving end-to-end joint search. This effectively solves the problems in neural architecture search where high-performance architectures are mistakenly eliminated due to the discrete and fixed evaluation policy, and the inability of evaluation parameters to adaptively adjust with architecture evolution.

[0013] 3. This invention proposes a hierarchical quality embedding technique for evaluation strategies based on hyperbolic Poincaré sphere projection. It aims to utilize the exponential capacity advantage of non-Euclidean geometric space to perform lossless modeling of evaluation strategies with tree-like hierarchical structures, and to accurately calculate the evaluation confidence through Riemann Gaussian distribution. This effectively solves the problems of severe distortion of policy semantic distance caused by traditional Euclidean space metrics and the lack of mathematical uncertainty quantification methods for evaluation results. Attached Figure Description

[0014] Figure 1 This is a diagram illustrating the overall architecture of the present invention; Figure 2 Build a complete flowchart for intelligent evaluation. Detailed Implementation

[0015] To make the objectives, content, and advantages of this invention clearer, the specific embodiments of this invention will be described in further detail below with reference to the accompanying drawings and examples. This invention relates to an intelligent evaluation method and apparatus for target detection models based on adaptive learning. It proposes a method for constructing a causal knowledge graph and transferring reasoning for evaluation strategies based on a structural causal model (SCM). This method utilizes a counterfactual intervention mechanism to generate virtual scenes at the feature level, effectively solving the problem that traditional evaluation is susceptible to interference from spurious environmental relevance, leading to policy failure in new scenes. Furthermore, it proposes a differentiable measurement method for architecture-policy coupling based on multi-view fusion, aiming to achieve end-to-end joint optimization of evaluation strategy parameters and model architecture through meta-weighted networks and differentiable relaxation techniques. This method utilizes a hyperbolic... The evaluation strategy quality embedding technique of the sphere model leverages the exponential capacity advantage of non-Euclidean geometry to solve the problem of distortion in hierarchical strategy measurements, and accurately quantifies the uncertainty of evaluation results through Riemannian-Gaussian distribution. This invention, based on causal decoupling and hyperbolic embedding, aims to reduce the cost of manual parameter tuning and improve the reliability of evaluation results in complex environments, thereby obtaining a robust, adaptive, and highly accurate intelligent evaluation method.

[0016] This invention designs an intelligent evaluation device for object detection models based on adaptive learning. The overall architecture of the device is shown in the figure below. Figure 1 As shown. The device mainly includes: a causal transfer evaluation inference module, an architecture-policy coupling degree differentiable measurement module, and a hyperbolic space quality embedding module; 1. Causal Transfer Assessment and Reasoning Module The core function of this module is to construct the causal dependency relationship between scenarios and policies, generate counterfactual samples, and screen out robust policies that are resistant to environmental interference. This module consists of three sub-modules: The scenario-policy dependency graph construction submodule includes a causal relationship discovery unit and a graph node definition unit. The causal relationship discovery unit uses a PC algorithm to scan the dataset metadata, mine the conditional independence between the scenario environment, model features, and evaluation policy parameters, and construct a directed acyclic graph. The graph node definition unit defines environmental factors as confusion nodes and policy parameters as intervention nodes, clarifying the causal flow between variables.

[0017] The counterfactual virtual scene generation submodule includes an intervention feature generation unit and a virtual sample synthesis unit. The intervention feature generation unit introduces a do-calculus intervention quantifier to sever the association between environmental nodes and target nodes in the feature space, thereby perturbing the environmental feature vector. The virtual sample synthesis unit then generates unseen counterfactual virtual scene features based on the perturbed features to simulate OOD scenarios.

[0018] The invariant risk minimization constraint submodule includes an environmental risk calculation unit and a robust policy selection unit. The environmental risk calculation unit calculates the evaluation loss of candidate policies in real-world and counterfactual virtual scenarios, respectively; the robust policy selection unit introduces environmental invariance constraints to eliminate policies that perform too differently under different environmental interventions, retaining a subset of robust policies that focus on the essential characteristics of the target.

[0019] 2. Module for Differentiable Measurement of Architecture-Policy Coupling The core function of this module is to establish a deep coupling relationship between the evaluation strategy and the model architecture, and to achieve automated strategy search through differentiable optimization. This module consists of three sub-modules: The multidimensional orthogonal feature mapping submodule includes a coupled dimension calculation unit and a feature orthogonalization unit. The coupled dimension calculation unit decomposes the matching relationship between the model and the policy into four dimensions: performance response, information entropy, Jacobian sensitivity, and computational cost. The feature orthogonalization unit decouples the feature vectors of the above four dimensions to ensure that each dimension remains orthogonal and independent in the metric space, thus avoiding feature redundancy.

[0020] The scene-driven meta-weight generation submodule includes a scene feature extraction unit and an attention weight prediction unit. The scene feature extraction unit analyzes the distribution characteristics of the current data to be evaluated, such as target density and occlusion ratio; the attention weight prediction unit uses a lightweight meta-network to dynamically predict the importance weight vectors of the above four dimensions based on scene features.

[0021] The policy parameter differentiable search submodule includes a parameter relaxation unit and a joint gradient update unit. The parameter relaxation unit uses the Gumbel-Softmax function to continuously relax discrete evaluation policy parameters, such as the NMS threshold and Top-K. The joint gradient update unit constructs a backpropagation path from the weighted evaluation metrics to the policy parameters, enabling the evaluation policy to undergo joint gradient descent during the model architecture search process and converge to the optimal policy parameters.

[0022] The joint gradient update unit uses the continuous policy weight vector z (see Section 3.3) output by Gumbel-Softmax relaxation as a differentiable policy feature representation. It backpropagates the gradient of the evaluation index loss L with respect to z to the policy parameters through the chain rule, thereby constructing a complete differentiable gradient propagation path from the weighted evaluation index to the discrete policy parameters. It then performs gradient descent together with the model architecture parameters to converge to the optimal evaluation policy configuration under the current architecture.

[0023] 3. Hyperbolic Space Quality Embedded Module The core function of this module is to solve the metric distortion problem of hierarchical strategies using non-Euclidean geometry and to quantify the confidence level of the evaluation results. This module consists of three sub-modules: The hyperbolic manifold projection submodule includes a curvature adaptive learning unit and an exponential mapping unit. The curvature adaptive learning unit automatically finds the optimal manifold curvature that fits the current policy tree level complexity through gradient descent. The exponential mapping unit uses the tangent space as an intermediary to project the Euclidean space-optimized policy feature vectors onto the hyperbolic manifold. On a spherical manifold, to preserve the inherent hierarchical semantic distance between strategies.

[0024] The policy quality decoupling encoding submodule includes a quality subspace partitioning unit and a manifold feature encoding unit. The quality subspace partitioning unit further decomposes the policy quality into four independent Riemannian subspaces—stability, discriminability, robustness, and efficiency—in hyperbolic space. The manifold feature encoding unit then encodes the policy vectors in each subspace to obtain a more refined quality representation.

[0025] The Riemann uncertainty quantification submodule includes a Riemann Gaussian distribution estimation unit and a confidence level determination unit. The Riemann Gaussian distribution estimation unit calculates the probability distribution of the strategy on the hyperbolic manifold, outputs the mean of the distribution as the final quality score, and calculates the geodesic variance of the distribution. The confidence level determination unit uses the geodesic variance as an uncertainty measure. When the uncertainty exceeds a preset threshold, it automatically marks the evaluation result as "low confidence" and triggers a manual review alarm.

[0026] Example 1: This invention proposes an adaptive evaluation method for target detection models based on causal decoupling and hyperbolic geometric embedding. The goal is to overcome the challenges of spurious environmental correlation and non-differentiable discrete evaluation strategies without human intervention, and to achieve accurate, robust, and quantifiable automatic evaluation of target detection models in open and dynamic scenarios.

[0027] 1. Overview of the Operation Process The intelligent evaluation system based on causal inference and hyperbolic geometry mainly includes the following four stages: Figure 2 As shown.

[0028] S1. Preparation and Causal Discovery Phase: This phase is the cornerstone of the entire system. Its main objective is to uncover hidden causal mechanisms from raw data and generate counterfactual augmented data capable of breaking spurious correlations. First, the system scans metadata to construct a structural causal model, defining the dependencies between environment, objective, and policy, and building a scenario-policy causal graph. Second, it uses the backdoor criterion to identify environmental confounding factors. Finally, by performing counterfactual interventions in the feature space, it generates a "counterfactual augmented dataset," providing an unbiased benchmark for subsequent decoupling evaluation. An Invariant Risk Minimization (IRM) mechanism is introduced.

[0029] S2. Differentiable Policy Search Phase: The core of this phase lies in breaking down the barrier between traditional discrete evaluation parameters and continuous model training. First, the model and policy are mapped to four orthogonal Riemannian manifold spaces: performance, information, sensitivity, and efficiency. Second, a meta-weight network is used to dynamically generate attention weights for each dimension based on scene features. Third, the Gumbel-Softmax relaxation technique is introduced to make the discrete policy parameters continuous. Finally, end-to-end joint gradient descent is performed through bi-layer programming to search for the optimal policy configuration under the current architecture.

[0030] S3, Hyperbolic Quality Embedding Stage: This stage aims to leverage the exponential capacity advantage of non-Euclidean geometry to address the metric distortion problem of complex hierarchical policies. The system uses exponential mapping to project policy vectors from Euclidean space onto a hyperbolic geometry with constant negative curvature. Inside the spherical manifold; the manifold curvature is automatically learned through gradient descent to adapt to the hierarchical structure of the policy; finally, the policy is modeled as a Riemannian Gaussian distribution, and the quality score and uncertainty confidence are output.

[0031] S4. Evaluation, Feedback, and Decision-Making Stage: This stage is responsible for translating the mathematical evaluation results into engineering decisions. Based on the uncertainty of the Riemann variance determination, it decides whether to adopt the automatic evaluation results or trigger manual review; at the same time, it feeds back high-confidence difficult samples to the model training loop, completing the "evaluation-feedback-optimization" closed loop.

[0032] 2. Preparation and Causal Discovery Stage This stage forms the cornerstone of the entire intelligent evaluation methodology, with its core task being to establish a causal understanding between data and policies. The main objectives include: First, constructing a high-fidelity scenario-policy causal graph to accurately describe the dependency topology between environmental factors, target features, evaluation policies, and model performance; Second, generating counterfactual intervention samples based on this graph to sever spurious statistical correlations between environmental background and targets at the feature level, providing an unbiased data foundation for subsequent robustness evaluation.

[0033] 2.1 Construction of Scenario-Strategy Causal Graph Traditional evaluation methods often rely on the assumption of independent and identically distributed (ICD) paths, neglecting the pervasive confounding biases in the real world. To address this issue, it is essential to first construct causal graphs to identify and block these interfering paths.

[0034] Structural Causal Model (SCM) Definition: This invention first defines a four-tuple structural causal model. The graph construction algorithm formally describes the causal mechanism of the evaluation system. Here, U is the set of exogenous variables, representing unobserved potential noise or background factors. V is the set of endogenous variables, i.e., the observable nodes in the graph. F is the set of structural equations, describing the generation mechanism between variables. P(U) is the probability distribution of the exogenous variables. The graph construction algorithm steps are as follows: Variable node definition: Defines the set of endogenous variables. Where E represents environmental feature nodes, including: illumination intensity, weather type, background texture complexity, etc. O represents image target feature nodes, including: object category semantics, bounding box position, scale, occlusion ratio, etc. The evaluation strategy parameter nodes include NMS threshold, confidence threshold, and Anchor aspect ratio configuration. The evaluation metrics include mAP, Recall, and Inference Time.

[0035] Dependency discovery: Conditional independence is tested using the PC algorithm (Peter-Clark Algorithm). Given Z... , then delete The edges between them.

[0036] Note: Here, X and Y are any two variable nodes to be tested in the PC algorithm, and do not specifically refer to the target feature node O or the evaluation index node defined above. Z is any subset of observed variables other than X and Y, used as the condition set. (The previous section discussed evaluation index nodes.) The meaning of Y here is different—the former is a specific definition node in the causal graph, while the latter is a generalization symbol in the conditional independence test.

[0037] Causal direction determination: Based on the V-structure and orientation rules, the direction of the edges is determined to form a directed acyclic graph (DAG).

[0038] Confusion factor identification: Identifying nodes that simultaneously point to the target feature node O and the evaluation metric node. The set of common parent nodes (i.e., the confusion factor C) satisfies and This set is the object that will require subsequent intervention and decoupling.

[0039] Note: A common parent node refers to a node in a directed acyclic graph (DAG) that has directed edges pointing to both the target feature node O and the evaluation index node. The nodes. In this framework, the typical confusion factor C is the environmental feature node E, that is, E->O and E-> The set of confounding factors C is the structural causal model M=<U,V,F,P(U)> A subset of the endogenous variable set V in O, whose existence will occur in O and The spurious statistical correlations between them need to be blocked by do-calculus interference.

[0040] 2.2 Counterfactual intervention and enhancement To eliminate the influence of the identified confounding factor C during the evaluation phase, this invention introduces an interference estimator in causal inference. It simulates a virtual scenario where "the physical environment changes but the objects remain the same" by generating counterfactual samples.

[0041] The specific implementation steps are as follows: Step S101: Construction of Decoupled Representation Learning Network Build and train a based Or a decoupled encoder based on the StyleGAN architecture Map the input image x to a decoupled latent feature space:

[0042] in, This is the environmental style factor (corresponding to E in the graph, capturing non-semantic information such as background, weather, and lighting); For the target content factor (corresponding to O in the atlas), semantic information such as object shape, texture, and category is captured. Loss is minimized through mutual information. constraint and Orthogonality.

[0043] Step S102: Causal Intervention Generation (Do-Intervention) Perform intervention operations in the latent feature space. Preserve the target content factors. Frozen and unchanged, for environmental style factors Perform intervention operations .

[0044] Intervention strategies: It simulates combinations that rarely occur in the real world by randomly sampling from the prior distribution N(0, I) or by shuffling environmental factors in the training set.

[0045] Step S103: Synthesis of Counterfactual Feature Map Using pre-trained generative decoder Counterfactual images or feature maps are generated based on the combination of features after intervention.

[0046] Should This refers to counterfactual samples that retain the target causal characteristics but remove the original environmental biases.

[0047] 2.3 Constraint on Minimizing Invariant Risk To ensure the robustness of the evaluation strategy, this invention introduces an Invariant Risk Minimization (IRM) mechanism during the strategy search phase. The following optimization objective is constructed:

[0048] Note: λ is the invariant risk regularization coefficient (hyperparameter, λ>0), used to balance two optimization objectives: the first term The first term is the sum of empirical risks across all training environments; the second term is the environment invariance penalty, and the forced feature indicates that the local optimum condition is satisfied in all environments; the larger λ is, the stronger the constraint on cross-environment invariance, and the less the model depends on spurious environment-related features.

[0049] Formula Explanation: These are the model and strategy parameters; It is a collection that includes the original environment and the counterfactual intervention environment; In a specific environment The assessment risk (Loss); For feature extractors.

[0050] illustrate: This is a neural feature extractor with parameter θ that maps the model input to an intermediate representation space. It extracts invariant features that have a stable causal relationship with target detection quality, while masking environmentally confusing features. IRM constraints are applied to the output of this feature extractor, forcing... The learned features maintain local optimum of prediction in all training and counterfactual intervention environments.

[0051] The second penalty term mandates: the optimal classifier. In all environments The results are all locally optimal. This means that the model must use stable causal features (objective O) for prediction, rather than unstable environmental features (environment E), thus mathematically eliminating spurious correlations.

[0052] Note: IRM constraints do not directly modify the image during the inference phase, but rather affect the learning objective during the training phase. Specifically: by using the counterfactual samples generated in step S103... Add to training environment set This forces the feature extractor It learns the same stable target characteristics when facing both the original environment and the counterfactual intervention environment; in conjunction with the environment invariance penalty term, it mathematically severs the connection between the environment E and the evaluation index node. The spurious correlation path between them ensures that the evaluation results are ultimately determined solely by the true characteristics of target O, rather than by environmental background interference.

[0053] The meaning of the second penalty formula: Let w be the penalty applied to the feature... scalar classifier weights on (θ), Let w be the gradient of the risk with respect to w at w=1; the penalty term requires that this gradient be close to zero in all training environments e (i.e., its L2 norm tends to 0). This is equivalent to: w=1 being the locally optimal classifier weight in all environments, meaning that the features (θ) itself can directly support optimal prediction in various environments, forcing the model to rely only on causal target features that are stable across environments for judgment.

[0054] 3. Differentiable Strategy Search Phase This stage addresses the problem that evaluation strategy parameters (such as Top-K and IOU thresholds) are usually discrete and non-differentiable, making it impossible to jointly optimize them with the model architecture.

[0055] 3.1 Multi-view orthogonal feature mapping Model architecture A and evaluation strategy The coupling relationship is decomposed into four orthogonal metric manifolds.

[0056] Dimension 1, Performance Response Manifold: Calculate the IOU distribution between the predicted bounding box and the true bounding box, using the Smooth-L1 Loss approximation of non-differentiable mAP.

[0057] Dimension 2, Information Entropy Manifold: Calculates the mutual information between the feature layer and the output layer to measure the degree of information loss.

[0058] Dimension 3, Jacobi Sensitivity: Calculated Output Out For input In Jacobian matrix norm The robustness of the quantification model to small perturbations.

[0059] Note: Here In Specifically refers to the input feature tensor of the model (such as feature maps or image pixels). Out Specifically, the detection output tensors of the model (such as the coordinates of the prediction box and the class confidence score) are all specific tensor variables.

[0060] Note: The subscript F denotes the Frobenius norm of the matrix, which is the standard mathematical notation for measuring the overall magnitude of the matrix, and is related to the structural causal model M=<U,V,F,P(U)> The meaning of F (the set of structural equations) in this context is completely different; it is merely a general symbol for the matrix norm.

[0061] Dimension 4, Hardware Efficiency Manifold: Construct a lookup-based latency and memory differentiable proxy model to map discrete hardware metrics to continuous functions.

[0062] 3.2 Scenario-Driven Meta-Weight Network To achieve "location-specific" evaluation, a lightweight meta-weight network is designed.

[0063] Input: Statistical distribution characteristics of the current scene (e.g., target density histogram, occlusion rate distribution).

[0064] Processing: Two fully connected layers (MLP) and the Softmax activation function.

[0065] Output: A dynamic weight vector with four dimensions. This means that in low-computing-power edge scenarios, the network will automatically increase δ (efficiency weight); in high-sensitivity security scenarios, it will automatically increase α (performance weight).

[0066] 3.3 The Gumbel-Softmax relaxation strategy employs a continuous relaxation method using the Gumbel-Softmax distribution for discrete evaluation parameters (e.g., NMS method selection: Greedy-NMS vs. Soft-NMS). Let the set of discrete strategies be... Its selection probability is Introducing reparameterization techniques: Note: Here ,..., For the set of candidate solutions for discrete evaluation strategies, each This corresponds to a complete evaluation strategy parameter configuration, and is consistent with the evaluation strategy parameter node mentioned earlier. The meaning is consistent. Gumbel-Softmax selects the probability of k candidate solutions. By performing continuous relaxation, the discrete policy selection process is transformed into a differentiable operation, allowing the evaluation policy to be jointly optimized through gradient descent.

[0067]

[0068] in For noise, This is the temperature coefficient. When... hour, Approximating a discrete one-hot vector; when hour, It is continuously differentiable, thus establishing a gradient propagation path from evaluation metrics to policy parameters. .

[0069] Meaning of each symbol: The continuous selection weights for the i-th policy candidate (taking values ​​(0,1) and... (), is a continuous approximation of discrete One-hot selection; For strategy The prior choice probability; ~Gumbel(0,1) represents independent and identically distributed Gumbel noise, introducing randomness to explore diverse strategies; τ is the temperature coefficient (when τ→0, z degenerates into a one-hot discrete choice, and when τ>0, z is continuously differentiable). To evaluate the gradient of the loss L with respect to the policy parameters, by The continuous differentiability of the gradient ensures that it can be backpropagated through the automatic differentiation framework, thereby incorporating discrete policy selection into end-to-end joint gradient optimization.

[0070] 4. Hyperbolic quality embedding stage This stage utilizes hyperbolic geometry to address the problem of measurement distortion in Euclidean space for hierarchical strategies and quantifies the uncertainty of the evaluation results.

[0071] 4.1 Poincaré spherical projection Evaluation strategy space typically presents a tree-like hierarchical structure (global strategy). Subclass strategy (Specific parameters). This embodiment utilizes the Poincaré Ball model. Embed it.

[0072] Step S301: Tangent Space Mapping. First, obtain the eigenvectors of the optimal policy in the tangent space (i.e., the Euclidean space at the origin). .

[0073] Parameter meaning: The policy feature vector in the Euclidean tangent space is generated by the linear layer of the policy encoder. The tangent space of the Poincaré sphere at the origin O is equivalent to the n-dimensional Euclidean space R. n ; Let be an n-dimensional Poincaré sphere with curvature parameter c, defined as {x∈R} n : c||x|| 2 <1}, where c>0 is the learnable manifold curvature.

[0074] Explanation: (1) Policy feature vector Obtaining: The continuous policy weights output by Gumbel-Softmax in Section 3.3 By mapping to a fixed-dimensional Euclidean vector through a linear projection layer, the strategy is transformed from probabilistic weights into computable geometric vectors. (2) Tangent space Meaning: Tangent space is a flat Euclidean space that is tangent to the manifold at a point; tangent space at the origin O of the Poincaré sphere. Equivalent to standard Euclidean space R n Gradient descent optimization can be performed directly. Being in tangent space means that it has not yet been projected and is still in flat space. (3) The meaning of the origin: Poincaré sphere The geometric center of the sphere corresponds to the top-level global root policy in the policy hierarchy and is the reference point for the exponential mapping.

[0075] Step S302: Exponential Mapping. Project the vector onto the hyperbolic manifold using the exponential mapping formula:

[0076] Parameter meaning: Let v be the exponential mapping at the origin of the Poincaré sphere, projecting the Euclidean vector v in the tangent space onto the hyperbolic manifold; c is the curvature parameter of the manifold; ||v|| is the L2 norm of vector v; The Euclidean length is compressed to the (0,1) interval to ensure that the modulus of the mapping result is strictly less than 1 / sqrt(c); is the unit direction vector of v; overall meaning: embedding geodesic distance compressed by tanh along the v direction on the hyperbolic manifold, naturally mapping Euclidean features to the hyperbolic hierarchical structure.

[0077] In this space, the root node strategy is located at the center of the sphere, and the subdivided leaf strategy is located at the edge, with the distance increasing exponentially, perfectly matching the tree structure.

[0078] Note: This space is an n-dimensional Poincaré sphere. That is, a hyperbolic Riemannian manifold with constant negative curvature can be regarded as satisfying c||x|| 2 An n-dimensional open sphere with a resolution of less than 1. Its core geometric property is that the geodesic distance between adjacent points increases faster closer to the sphere's boundary, naturally fitting a hierarchical strategy tree structure—the root node is located at the sphere's center, and the finer-grained strategies at lower levels are farther from the sphere's center. The exponential growth of the distance between nodes accurately reflects the degree of strategy refinement, thus avoiding the distortion of hierarchical measurements in Euclidean space.

[0079] Step S303: Curvature adaptive learning will adjust the manifold curvature. c Set as a learnable parameter. Minimize embedding distortion loss through gradient descent and automatically find the optimal curvature that fits the complexity of the current policy tree.

[0080] Implementation details: Initialize the curvature c to a positive scalar parameter (e.g., c=1.0) and add it to the model's learnable parameter list. Define the embedding distortion as:

[0081] in, For Poincaré and geodesic distance, This is the shortest path distance between two nodes i and j in the policy tree. It is calculated during each training round. The standard Adam optimizer updates c, which automatically adapts the degree of manifold bending to the complexity of the current policy hierarchy tree. As the policy hierarchy becomes deeper and the structure becomes more complex, c automatically increases to provide stronger hierarchy separation capabilities.

[0082] 4.2 Quantification of Uncertainty in Riemann-Gaussian Distribution To quantify the reliability of the evaluation, this embodiment does not output a single score, but instead outputs a Riemannian Gaussian distribution defined on hyperbolic space. The probability density function of the Riemannian Gaussian distribution is defined as: Note: The Riemannian Gaussian distribution here is the probability distribution of policy quality in hyperbolic space, describing the probability density of the quality score of each policy point z on the Poincaré sphere. Its form is similar to the Euclidean Gaussian distribution, but the Euclidean distance is replaced by hyperbolic geodesic distance. (z,μ) makes the distribution shape conform to the geometry of hyperbolic space, and can accurately model the quality uncertainty in a set of evaluation strategies with hierarchical structure.

[0083]

[0084] Parameter meaning: Lowercase z represents the Poincaré ball. Any policy point is the integral variable of the probability density function; uppercase Z(σ) is the normalization constant, ensuring that the integral of the probability density over the entire Poincaré sphere equals 1; μ is the distribution mean, i.e., the geometric center of the optimal policy on the Poincaré sphere, which serves as a point estimate for the final quality score; σ is the geodesic standard deviation, which quantifies the uncertainty of the evaluation results. (z,μ) represents the hyperbolic geodesic distance between z and μ.

[0085] mean : Represents the geometric center of the optimal evaluation strategy on the Poincaré sphere, serving as the final quality score.

[0086] Geodesic variance : Represents the degree of dispersion of the evaluation results, serving as a measure of uncertainty.

[0087] Decision logic: Set threshold .like This indicates that the model's performance is extremely unstable in the current counterfactual scenario, and the evaluation conclusions are unreliable.

[0088] 5. Evaluation, Feedback, and Decision-Making Stage Finally, all proven and effective evaluation strategies and scenario characteristics will be aggregated.

[0089] 5.1 Dynamic updates to the strategy library The SimHash algorithm is used to calculate the fingerprint of the newly generated counterfactual scene description, which is then compared with the historical scene database to remove duplicates. For newly admitted scene-policy pairs, their Riemann centroids on the Poincaré sphere are calculated, and the global evaluation benchmark is updated.

[0090] Specific implementation: (1) Calculate the Riemann centroid: Use the Riemann gradient descent iterative method—starting from the initial point p0, repeatedly execute: grad , until convergence, yielding the Riemann center of mass μ*; (2) Update the global baseline: merge the old and new centroids in hyperbolic space using an exponential moving average: α is the update learning rate, which ensures smooth and gradual updates of the global benchmark on the Poincaré sphere.

[0091] Note: The Riemannian centroid is a generalization of the Euclidean centroid to Riemannian manifolds, defined as the point that minimizes the sum of squared geodesic distances to all given points. On the Poincaré sphere, the Riemann centroid represents the geometric center of a set of scenario-policy pairs in hyperbolic space, reflecting the overall distribution of these policies and used to continuously update the global evaluation benchmark.

[0092] A candidate point on the Poincaré sphere, whose value is limited to the Poincaré sphere space. Inside; For the first Embedded representations of scene-policy pairs on the Poincaré sphere; Poincaré's goal and Geodesic distance between them; This represents the sum of the squared geodesic distances from the candidate point to all scene-policy pair embedding points. This formula finds the geometric center point that reflects the overall distribution of these scene-policy pairs in hyperbolic space by minimizing this sum of squared distances.

[0093] The formula as a whole operates on a Poincaré sphere (hyperbolic space). and They are a pair of inverse Riemannian mappings, the former "pulling" points on the manifold into the tangent space for computation, and the latter "pushing" the results in the tangent space back onto the manifold. It is the geodesic distance in hyperbolic space, replacing Euclidean distance to measure the distance between points. It is the Riemann centroid that minimizes the sum of the squares of the geodesic distances of all data points, obtained by using... The gradient descent step size is obtained through iteration. Then control the new centroid Integrate into global benchmark The speed enables smooth and gradual updates of the benchmark.

[0094] 5.2 Closed-loop optimization feeds back high-risk and high-uncertainty scenarios discovered during the evaluation process to the model training end as difficult samples, thereby improving the robustness of the object detection model in the next iteration.

[0095] The beneficial effects of this invention are: This invention provides an adaptive evaluation construction method and apparatus for target detection models based on causal decoupling and hyperbolic geometry embedding, which has the following beneficial effects: 1. This invention proposes a causal transfer evaluation and reasoning technique based on scenario-policy dependency graph. It aims to eliminate spurious correlations between environmental background and target at the feature level by using a structural causal model and a counterfactual intervention mechanism. This effectively solves the technical problem of traditional evaluation methods failing to evaluate models in new cross-domain scenarios and having inaccurate generalization ability assessments due to biased training data distribution.

[0096] 2. This invention proposes a method for measuring the differentiable degree of architecture-policy coupling based on multi-view orthogonal feature fusion. It aims to establish a fully differential gradient link between evaluation policy parameters and model architecture by using meta-weighted networks and Gumbel-Softmax relaxation techniques, thereby achieving end-to-end joint search. This effectively solves the problems in neural architecture search where high-performance architectures are mistakenly eliminated due to the discrete and fixed evaluation policy, and the inability of evaluation parameters to adaptively adjust with architecture evolution.

[0097] 3. This invention proposes a hierarchical quality embedding technique for evaluation strategies based on hyperbolic Poincaré sphere projection. It aims to utilize the exponential capacity advantage of non-Euclidean geometric space to perform lossless modeling of evaluation strategies with tree-like hierarchical structures, and to accurately calculate the evaluation confidence through Riemann Gaussian distribution. This effectively solves the problems of severe distortion of policy semantic distance caused by traditional Euclidean space metrics and the lack of mathematical uncertainty quantification methods for evaluation results.

[0098] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A causal decoupling adaptive evaluation device for a target detection model, characterized in that, The device includes: a causal transfer evaluation inference module, an architecture-policy coupling degree differentiable measurement module, and a hyperbolic space quality embedding module; The causal transfer evaluation and reasoning module is used to construct the causal dependency relationship between scenarios and policies, generate counterfactual samples, and screen out robust policies resistant to environmental interference; it includes three sub-modules: The scenario-policy dependency graph construction submodule includes a causal relationship discovery unit and a graph node definition unit. The causal relationship discovery unit uses the PC algorithm to scan the dataset metadata, mine the conditional independence between the scenario environment, model features and evaluation policy parameters, and construct a directed acyclic graph. The graph node definition unit defines environmental factors as confusion nodes and policy parameters as intervention nodes, clarifying the causal flow between variables. The counterfactual virtual scene generation submodule includes an intervention feature generation unit and a virtual sample synthesis unit. The intervention feature generation unit introduces a do-calculus intervention quantifier to sever the association between environmental nodes and target nodes in the feature space and apply perturbation to the environmental feature vector. The virtual sample synthesis unit generates unseen counterfactual virtual scene features based on the perturbed features to simulate OOD scenarios. The invariant risk minimization constraint submodule includes an environmental risk calculation unit and a robust policy selection unit. The environmental risk calculation unit calculates the evaluation loss of candidate policies in real scenarios and counterfactual virtual scenarios, respectively. The robust policy selection unit introduces environmental invariance constraints to eliminate policies that perform too differently under different environmental interventions and retains a subset of robust policies that focus on the essential characteristics of the target. The architecture-policy coupling degree differentiable measurement module is used to establish a deep coupling relationship between the evaluation policy and the model architecture, and to achieve automated policy search through differentiable optimization; it includes three sub-modules: The multidimensional orthogonal feature mapping submodule includes a coupled dimension calculation unit and a feature orthogonalization unit. The coupled dimension calculation unit decomposes the matching relationship between the model and the policy into four dimensions: performance response, information entropy, Jacobi sensitivity, and computational cost. The feature orthogonalization unit decouples the feature vectors of the above four dimensions to ensure that each dimension remains orthogonal and independent in the metric space. The scene-driven meta-weight generation submodule includes a scene feature extraction unit and an attention weight prediction unit. The scene feature extraction unit analyzes the distribution characteristics of the current data to be evaluated. The attention weight prediction unit uses a lightweight meta-network to dynamically predict the importance weight vectors of the above four dimensions based on the scene features. The policy parameter differentiable search submodule includes a parameter relaxation processing unit and a joint gradient update unit. The parameter relaxation processing unit uses the Gumbel-Softmax function to continuously relax the discrete evaluation policy parameters. The joint gradient update unit constructs a backpropagation path from the weighted evaluation index to the policy parameters, so that the evaluation policy can undergo joint gradient descent during the search process of the model architecture and converge to the optimal policy parameters. The hyperbolic space quality embedding module is used to address the metric distortion problem of hierarchical strategies using non-Euclidean geometry and to quantify the confidence of evaluation results; it includes three sub-modules: The hyperbolic manifold projection submodule includes a curvature adaptive learning unit and an exponential mapping unit. The curvature adaptive learning unit automatically finds the optimal manifold curvature that fits the current policy tree level complexity through gradient descent. The exponential mapping unit uses the tangent space as an intermediary to project the policy feature vector optimized in Euclidean space onto the hyperbolic manifold. On a spherical manifold, to preserve the inherent hierarchical semantic distance between policies; The policy quality decoupling encoding submodule includes a quality subspace partitioning unit and a manifold feature encoding unit. The quality subspace partitioning unit further decomposes the policy quality into four independent Riemann subspaces in hyperbolic space: stability, discriminability, robustness, and efficiency. The manifold feature encoding unit encodes the policy vector in each subspace to obtain a more refined quality representation. The Riemann uncertainty quantification submodule includes a Riemann Gaussian distribution estimation unit and a confidence level determination unit. The Riemann Gaussian distribution estimation unit calculates the probability distribution of the strategy on the hyperbolic manifold, outputs the mean of the distribution as the final quality score, and calculates the geodesic variance of the distribution. The confidence level determination unit uses the geodesic variance as an uncertainty measure. When the uncertainty exceeds a preset threshold, it automatically marks the evaluation result as "low confidence" and triggers a manual review alarm.

2. The target detection model causal decoupling adaptive evaluation device as described in claim 1, characterized in that, The joint gradient update unit uses the continuous policy weight vector z output by Gumbel-Softmax relaxation as a differentiable policy feature representation. It backpropagates the gradient of the evaluation index loss L with respect to z to the policy parameters through the chain rule, thereby constructing a complete differentiable gradient propagation path from the weighted evaluation index to the discrete policy parameters. It then performs gradient descent together with the model architecture parameters to converge to the optimal evaluation policy configuration under the current architecture.

3. An adaptive evaluation method for causal decoupling of target detection models, characterized in that, The method includes the following steps: S1. Preparation and Causal Discovery Phase: The system scans metadata to construct a structural causal model, defines the dependencies between environment, objectives, and policies, and builds a scenario-policy causal graph; secondly, it uses the backdoor criterion to identify environmental confounding factors; finally, it generates a "counterfactual augmentation dataset" by performing counterfactual interventions in the feature space; and introduces an invariant risk minimization (IRM) mechanism. S2. Differentiable Policy Search Stage: First, the model and policy are mapped to four orthogonal Riemannian manifold spaces: performance, information, sensitivity, and efficiency. Second, the meta-weight network is used to dynamically generate attention weights for each dimension based on scene features. Third, the Gumbel-Softmax relaxation technique is introduced to make the discrete policy parameters continuous. Finally, end-to-end joint gradient descent is performed through bi-layer programming to search for the optimal policy configuration under the current architecture. S3, Hyperbolic Quality Embedding Stage: Projecting the policy vector in Euclidean space onto a hyperbolic space with constant negative curvature using an exponential mapping. Inside the spherical manifold; the manifold curvature is automatically learned through gradient descent to adapt to the hierarchical structure of the policy; finally, the policy is modeled as a Riemannian Gaussian distribution, and the quality score and uncertainty confidence are output. S4. Evaluation, Feedback and Decision-Making Stage: Transform mathematical evaluation results into engineering decisions; based on the uncertainty of Riemann variance determination, decide whether to adopt automatic evaluation results or trigger manual review; at the same time, feed back high-confidence difficult samples to the model training loop to complete the "evaluation-feedback-optimization" closed loop.

4. The target detection model causal decoupling adaptive evaluation device as described in claim 3, characterized in that, In S1, the system scans metadata to construct a structural causal model, defines the dependencies between environment, target and policy, and constructs a scenario-policy causal graph. Secondly, using backdoor criteria to identify environmental confounding factors includes: Structural Causal Model (SCM) Definition: First, define a four-tuple structural causal model. The graph construction algorithm formally describes the causal mechanism of the evaluation system; where U is the set of exogenous variables, representing unobserved potential noise or background factors; V is the set of endogenous variables, i.e., the observable nodes in the graph; F is the set of structural equations, describing the generation mechanism between variables; P(U) is the probability distribution of exogenous variables; the graph construction algorithm steps are as follows: Define the set of endogenous variables Where E is the environmental feature node, including: illumination intensity, weather type and background texture complexity; O is the image target feature node, including: object category semantics, bounding box position, scale and occlusion ratio; The evaluation strategy parameter nodes include NMS threshold, confidence threshold, and Anchor aspect ratio configuration; As evaluation indicator nodes; Dependency discovery: Conditional independence is tested using the PC algorithm; given Z... , then delete The edges between; X and Y are any two variable nodes to be tested in the PC algorithm, and Z is the condition set; Causal direction determination: Based on the V-structure and orientation rules, the direction of the edges is determined to form a directed acyclic graph (DAG); Confusion factor identification: Identifying nodes that simultaneously point to the target feature node O and the evaluation metric node. The set of common parent nodes, i.e., the confusion factor C, satisfies and This set is the object that needs to be decoupled and intervened upon later.

5. The target detection model causal decoupling adaptive evaluation device as described in claim 4, characterized in that, In S1, generating a "counterfactual augmentation dataset" by performing counterfactual intervention in the feature space includes: Introducing the interferometer in causal inference It simulates a virtual scenario where "the physical environment changes but the objects remain unchanged" by generating counterfactual samples; Step S101: Construction of Decoupled Representation Learning Network Build and train a based Or a decoupled encoder based on the StyleGAN architecture Map the input image x to a decoupled latent feature space: in, Environmental style factors; The target content factor is used; the loss is minimized through mutual information. constraint and Orthogonality; Step S102, Causal Intervention Generation Perform intervention operations in the latent feature space; preserve the target content factors. Frozen and unchanged, for environmental style factors Perform intervention operations ; Intervention strategies: It is obtained by randomly sampling from the prior distribution N(0, I) or by shuffling the environmental factors in the training set, thereby simulating combinations that rarely occur in the real world; Step S103: Synthesis of Counterfactual Feature Map Using pre-trained generative decoder Counterfactual images or feature maps are generated based on the combination of features after intervention. Should This refers to counterfactual samples that retain the target causal characteristics but remove the original environmental biases.

6. The target detection model causal decoupling adaptive evaluation device as described in claim 5, characterized in that, In S1, the introduction of the invariant risk minimization (IRM) mechanism includes: The following optimization objective is proposed: Note: λ is the constant risk regularization coefficient, used to balance two optimization objectives: the first term The first term is the sum of empirical risks across all training environments; the second term is the environment invariance penalty, and the forced feature indicates that the local optimum condition is satisfied in all environments. These are the model and strategy parameters; It is a collection that includes the original environment and the counterfactual intervention environment; In a specific environment The following evaluation risks; For feature extractors; optimal classifiers In all environments The results are all locally optimal; this means that the model must use stable causal features for prediction, rather than unstable environmental features.

7. The target detection model causal decoupling adaptive evaluation device as described in claim 3, characterized in that, In step S2, the model and policy are mapped to four orthogonal Riemannian manifold spaces: performance, information, sensitivity, and efficiency. Secondly, a meta-weighted network is used to dynamically generate attention weights for each dimension based on scene features, including: Model architecture A and evaluation strategy The coupling relationship is decomposed into four orthogonal metric manifolds; Dimension 1, Performance Response Manifold: Calculate the IOU distribution between the predicted bounding box and the truth box; Dimension 2, Information Entropy Manifold: Calculates the mutual information between the feature layer and the output layer to measure the degree of information loss; Dimension 3, Jacobi Sensitivity: Calculated Output Out For input In Jacobian matrix norm The robustness of the quantification model to small perturbations; the subscript F denotes the Frobenius norm of the matrix; Dimension 4, Hardware Efficiency Manifold: Construct a lookup-based latency and memory differentiable proxy model to map discrete hardware metrics into continuous functions; Design a lightweight meta-weighted network: Input: Statistical distribution characteristics of the current scene ; Processing: Using two fully connected MLP layers and the Softmax activation function; Output: A dynamic weight vector with four dimensions. .

8. The target detection model causal decoupling adaptive evaluation device as described in claim 7, characterized in that, In S2, the Gumbel-Softmax relaxation technique is introduced to make the discrete strategy parameters continuous; Finally, by performing end-to-end joint gradient descent through bi-level programming, the optimal policy configuration under the current architecture is searched, including: For discrete evaluation parameters, a continuous relaxation is performed using the Gumbel-Softmax distribution; let the set of discrete policies be... Its selection probability is Introducing reparameterization techniques: ,..., For the set of candidate solutions for discrete evaluation strategies, each This corresponds to a complete evaluation strategy parameter configuration; the selection probability of Gumbel-Softmax among k candidate solutions. By performing continuous relaxation, the discrete policy selection process is transformed into a differentiable operation, allowing the evaluation policy to be jointly optimized through gradient descent. in For noise, For temperature coefficient; when hour, Approximating a discrete one-hot vector; when hour, It is continuously differentiable, thus establishing a gradient propagation path from evaluation metrics to policy parameters. ; Let be the continuous selection weight for the i-th policy candidate, taking values ​​(0,1), and It is a continuous approximation of discrete One-hot selection; For strategy The prior choice probability; ~Gumbel(0,1) represents independent and identically distributed Gumbel noise, introducing randomness to explore diverse strategies; τ is the temperature coefficient. To evaluate the gradient of the loss L with respect to the policy parameters, by The continuous differentiability of the gradient ensures that it can be backpropagated through the automatic differentiation framework, thereby incorporating discrete policy selection into end-to-end joint gradient optimization.

9. The target detection model causal decoupling adaptive evaluation device as described in claim 3, characterized in that, In S3, the policy vector in Euclidean space is projected onto the interior of a Poincaré spherical manifold with constant negative curvature using an exponential mapping; the hierarchical structure for automatically learning the manifold curvature to adapt the policy through gradient descent includes: The evaluation strategy space presents a tree-like hierarchical structure, utilizing the Poincaré Ball model. Embed; Step S301: Tangent Space Mapping. First, obtain the feature vector of the optimal policy in the tangent space. ; The policy feature vector in the Euclidean tangent space is generated by the linear layer of the policy encoder. The tangent space of the Poincaré sphere at the origin O is equivalent to the n-dimensional Euclidean space R. n ; Let be an n-dimensional Poincaré sphere with curvature parameter c, defined as {x∈R} n : c||x|| 2 <1}, where c>0 is the learnable manifold curvature; Step S302, Exponential Mapping Projecting eigenvectors onto a hyperbolic manifold using the exponential mapping formula: Let v be the exponential mapping at the origin of the Poincaré sphere, projecting the Euclidean vector v in the tangent space onto the hyperbolic manifold; c is the curvature parameter of the manifold; ||v|| is the L2 norm of vector v; The Euclidean length is compressed to the (0,1) interval to ensure that the modulus of the mapping result is strictly less than 1 / sqrt(c); Let v be the unit direction vector; overall meaning: embedding along the v direction on the hyperbolic manifold using tanh-compressed geodesic distances, naturally mapping Euclidean features to the hyperbolic hierarchical structure; in this space, the root node strategy is located at the center of the sphere, the subdivision leaf strategy is located at the edge, and the distance grows exponentially, perfectly fitting the tree structure. Step S303, Curvature Adaptive Learning manifold curvature c Set as a learnable parameter; minimize embedding distortion through gradient descent and automatically find the optimal curvature that fits the complexity of the current policy tree; Specific implementation: Initialize curvature c as a positive scalar parameter and add it to the model's learnable parameter list; define the embedding distortion as: in, For Poincaré and geodesic distance, The shortest path distance between two nodes i and j in the policy tree; calculated during each training round. The standard Adam optimizer updates c, allowing the degree of manifold curvature to automatically adapt to the complexity of the current policy hierarchy tree.

10. The target detection model causal decoupling adaptive evaluation device as described in claim 9, characterized in that, In S3, the strategy is modeled as a Riemannian Gaussian distribution, and the output quality score and uncertainty confidence level include: The probability density function of the Riemann-Gaussian distribution is defined as: The lowercase 'z' represents the Poincaré ball. Any policy point is the integral variable of the probability density function; uppercase Z(σ) is the normalization constant, ensuring that the integral of the probability density over the entire Poincaré sphere equals 1; μ is the distribution mean, i.e., the geometric center of the optimal policy on the Poincaré sphere, which serves as a point estimate for the final quality score; σ is the geodesic standard deviation, which quantifies the uncertainty of the evaluation results. (z,μ) represents the hyperbolic geodesic distance between z and μ; mean : Represents the geometric center of the optimal evaluation strategy on the Poincaré sphere, serving as the final quality score; Geodesic variance : Represents the degree of dispersion of the evaluation results, serving as a measure of uncertainty; Decision logic: Set threshold ;like This indicates that the model's performance is extremely unstable in the current counterfactual scenario, and the evaluation conclusions are unreliable.

Images