A reinforcement learning combined backbone module method and system for remote sensing images

Abstract

This invention discloses a method and system for reinforcement learning-based backbone modules for remote sensing images in the fields of remote sensing image processing and deep learning. The method includes the following steps: input preprocessing and construction of a heterogeneous primitive library; heterogeneous feature alignment and normalization representation; construction of a dynamic routing decision space and generation of routing actions; updating the reward network and routing strategy based on bi-layer primitive optimization; stabilizing policy updates based on reference policy regularization; and performing dynamic feature aggregation, task prediction, and controllable inference. This invention can adaptively allocate computational resources according to the complexity of remote sensing image samples, significantly reducing redundant computation while ensuring performance in detecting complex scenes, small targets, and dense targets. Furthermore, a single model can adapt to deployment scenarios with different computational constraints, making it widely applicable in engineering.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing image processing and deep learning technology, specifically to a method and system for combining reinforcement learning backbone modules for remote sensing images. Background Technology

[0002] With the rapid development of Earth observation technology, high-resolution remote sensing images have been widely used in fields such as surface monitoring, resource management, disaster assessment, and urban planning. Compared with natural scene images, high-resolution remote sensing images typically have characteristics such as large spatial range, uneven information distribution, wide target scale span, arbitrary target orientation, and complex background interference. A single image may contain large areas of low-information-density background regions as well as areas with dense and complex targets, posing a dual challenge to the feature extraction capabilities and computational efficiency of visual models.

[0003] Most existing deep learning-based visual recognition methods employ static deep networks as the backbone for feature extraction, such as Convolutional Neural Networks (CNNs), Visual Transformers (ViTs), or hybrid structures of both. These models, once trained, have a fixed network topology and computational flow, executing the same computational path for all input samples. This leads to two main problems: first, redundant computation occurs in large, simple background regions; second, in regions with densely packed small targets, drastic scale changes, or complex background textures, insufficient representation occurs due to the fixed model capacity or computational path, affecting detection or recognition accuracy.

[0004] To alleviate these problems, some studies have introduced dynamic neural network mechanisms, such as early exit, skip-layer execution, and token pruning. These methods can reduce average computation to some extent, but their dynamic decision rules are usually relatively simple, relying on preset thresholds or manually set path selection strategies, making it difficult to fully characterize the complex characteristics of high-resolution remote sensing scenes. Furthermore, some methods model the dynamic routing process as a Markov decision process and use reinforcement learning for optimization; however, the reward function is often a manually set static scalar form, unable to adapt to dynamic changes in the training phase, sample complexity, and computational constraints, easily leading to the routing strategy converging to a suboptimal path. In addition, existing dynamic networks have a fixed accuracy-efficiency tradeoff after training; when the computing power, power consumption, or latency constraints of the deployment environment change, retraining or additional fine-tuning is often required, making it difficult for a single model to achieve flexible adjustment during runtime.

[0005] Therefore, how to reduce redundant calculations while ensuring the accuracy of remote sensing image resolution, and how to achieve flexible adaptation to different deployment environments, are technical problems that urgently need to be solved in this field. Summary of the Invention

[0006] The purpose of this invention is to provide a reinforcement learning-based backbone module method and system for remote sensing images, which solves the technical problems of unreasonable allocation of static network resources, difficulty in aligning heterogeneous features in dynamic networks, poor stability of routing optimization, and insufficient deployment flexibility, thereby achieving adaptive and efficient parsing of high-resolution remote sensing images.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows: A reinforcement learning-based backbone module combination method for remote sensing images includes the following steps: S1. Input preprocessing and heterogeneous primitive library construction: Receive high-resolution remote sensing images and perform standardized preprocessing. Construct a heterogeneous feature extraction primitive library containing convolution primitives, self-attention primitives, multi-scale primitives, lightweight primitives and hybrid structure primitives. Configure unified input and output interfaces and computational cost attributes for each primitive. S2. Heterogeneous Feature Alignment and Standardized Representation: Through the feature topology alignment interface, the spatial scale, channel dimension and representation of the features output by different candidate primitives are transformed, converting convolutional spatial grid features and self-attention sequence features into a unified standardized feature representation; S3. Construct a dynamic routing decision space and generate routing actions: Construct a state representation based on current intermediate features, historical routing trajectories, network layer positions, candidate primitive states, consumed computing resources, and preference parameters; define candidate primitive selection, path switching, skipping layers, and early exit as action sets; the module-combined agent outputs the action probability distribution based on the state representation to generate a dynamic feature extraction path for the current input sample; S4. Update the reward network and routing strategy based on two-layer meta-optimization: In the outer layer optimization process, the meta-reward network generates an intrinsic reward signal based on the state-action trajectory, detection loss, computational cost, and preference parameters, and updates the meta-reward network parameters based on the comprehensive evaluation target of the validation set; In the inner layer optimization process, the module-combined agent updates the routing strategy parameters according to the intrinsic reward signal. S5. Stable policy update based on reference policy regularization: Introduce Kullback-Leibler divergence constraints between the current routing policy and the reference routing policy in the inner policy optimization objective to limit the difference in the output action probability distribution of the two under the same state; S6. Perform dynamic feature aggregation, task prediction and controllable inference: Perform scale unification, channel mapping, local detail enhancement and global context aggregation on the features output by different dynamic paths to obtain unified task features and input them into the adaptive detection head to complete task prediction; during the deployment phase, adjust the computation path according to preference parameters and budget constraints to achieve controllable inference at different accuracy and efficiency working points.

[0008] Furthermore, the feature topology alignment interface in S2 is used to perform forward projection of the spatial grid features output by the convolution primitive into serialized label features that can be processed by the self-attention primitive, or to reverse reconstruction of the serialized label features output by the self-attention primitive into spatial grid features that can be processed by the convolution primitive.

[0009] Furthermore, the state representation in S3 is generated by concatenating global semantic feature representation, historical module selection trajectory, deep hierarchical position embedding, and continuous preference label, followed by nonlinear mapping and self-attention enhancement; the continuous preference label is used to quantify the degree of preference for detection accuracy and computational cost, with a value range of [0,1] and the sum of the two is 1.

[0010] Furthermore, in S4, the objective of the outer layer optimization is to maximize the joint optimality of detection accuracy and computational efficiency on the validation set; the inner layer reinforcement learning uses the policy gradient method to update the routing policy parameters, and the objective function is to introduce KL divergence regularization based on the reference policy on the basis of traditional cumulative reward.

[0011] Furthermore, the reference routing strategy in S5 is determined by the strategy parameters of the previous round, the historical average strategy parameters, the exponential moving average strategy parameters, or the strategy parameters saved in the stable phase.

[0012] Furthermore, the dynamic feature aggregation process in S6 includes: The adaptive feature gateway maps the dynamic features output from any path to a preset standard feature subspace. Based on the feature response distribution, a spatial response map and channel weight vector are generated, and local detail enhancement branches and global context enhancement branches are constructed respectively. The fusion coefficients of the two branches are adaptively calculated based on the complexity of the sample content to obtain the final aggregated features.

[0013] Furthermore, the controllable reasoning process in S6 includes: The runtime control vector is input into the policy network. The control vector includes the preference for recognition accuracy, the degree of constraint on inference latency, the preference for energy consumption control, and the sensitivity to the occupation of video memory or on-chip cache. During the reasoning process, the remaining computational budget is continuously perceived, and action sampling is restricted to the budgetable region. It supports dividing remote sensing images into multiple local regions and assigning different control parameters according to the complexity of each region to achieve hierarchical and controllable inference.

[0014] A reinforcement learning ensemble backbone module system for remote sensing images, comprising: The heterogeneous primitive library construction module is used to build a heterogeneous feature extraction primitive library containing at least two types of feature extraction primitives, and to configure a unified input / output interface and computational cost attribute for each primitive. Dynamic routing decision module: contains a module combination agent, used to construct state representation and action space, and generate dynamic feature extraction paths; Two-layer meta-optimization module: includes a meta-reward network, used for outer-layer meta-reward learning and inner-layer routing strategy optimization; Policy stabilization module: Used to introduce reference policy KL divergence regularization to stabilize the policy update process; Controllable Inference and Flexible Deployment Module: Used to receive deployment preference parameters during the deployment phase and control the trained module combination agent to dynamically select feature extraction paths for the input image.

[0015] A reinforcement learning ensemble backbone module system for remote sensing images also includes: Heterogeneous Feature Alignment Module: Used to standardize the spatial, channel, and representational forms of heterogeneous features output by different primitives; The dynamic feature aggregation and adaptive detection head module is used to aggregate the output features of dynamic paths and complete target detection, orientation angle prediction or instance segmentation tasks through the adaptive detection head.

[0016] Furthermore, the adaptive detection head in the dynamic feature aggregation and task prediction module adopts a structural decoupling design, which includes parallel classification branches, directional regression branches, and optional segmentation branches, which output class probabilities, rotated bounding box parameters, and pixel-level masks, respectively.

[0017] By adopting the above technical solution, the present invention has the following advantages: 1. This invention provides a method and system for reinforcement learning to combine backbone modules for remote sensing images. By using reinforcement learning agents to dynamically generate network topology based on the target scale, density, and background complexity of input samples, it can reduce redundant computation in simple background areas and enhance feature representation in complex areas, significantly improving the adaptability to high-resolution remote sensing scenes and making the scene adaptability stronger.

[0018] 2. This invention provides a method and system for reinforcement learning assembling backbone modules for remote sensing images. It constructs an extensible primitive library containing multiple heterogeneous primitives and designs a unified feature alignment interface. It has strong heterogeneous feature fusion capabilities, effectively solves the granularity mismatch problem between CNN and ViT features, and can fully combine the advantages of local detail modeling and global context modeling.

[0019] 3. This invention provides a reinforcement learning-based backbone module method and system for remote sensing images. It adopts a bi-level optimization mechanism to adaptively generate intrinsic reward signals, resulting in more stable route optimization and eliminating the dependence on manually set static reward functions. The introduction of reference policy regularization constraints effectively suppresses policy oscillations and improves the stability of dynamic architecture search. Compared with static models with fixed computational processes, this invention can reduce the average computational load by 30%-50% while maintaining comparable or even better detection performance. It is particularly suitable for batch processing of large-scale remote sensing images with higher computational efficiency.

[0020] 4. This invention provides a reinforcement learning combination backbone module method and system for remote sensing images. It achieves continuous adjustment of runtime accuracy and efficiency through preference parameters. A single model can be adapted to deployment scenarios with different computing power constraints, such as ground data centers, spaceborne platforms, and airborne edge devices, without the need for retraining. It adopts a modular design, and the novel feature extraction structure can be directly connected to the dynamic routing framework as a candidate primitive without the need to redesign the entire network, which has good engineering reuse value. Attached Figure Description

[0021] Figure 1 This is an overall flowchart of the reinforcement learning ensemble backbone module method for remote sensing images according to the present invention; Figure 2 This is a schematic diagram illustrating the process of constructing a heterogeneous feature extraction primitive library and aligning data in an embodiment of the present invention. Figure 3 This is a schematic diagram illustrating the structure and interaction of the modular combined intelligent agent and the two-layer optimized timing decision module in an embodiment of the present invention; Figure 4 This is a flowchart illustrating the dynamic feature aggregation and adaptive detection head in an embodiment of the present invention; Figure 5 This is a schematic diagram of the process of controllable inference and flexible deployment in this invention. Detailed Implementation

[0022] The technical solution of the present invention will be specifically described below with reference to the accompanying drawings. It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0023] Example 1: A reinforcement learning method for combining backbone modules for remote sensing images

[0024] This invention provides a reinforcement learning-based backbone module combination method for remote sensing images, specifically as follows: Figure 1 As shown, the method of this invention includes the following six main steps: S1, input preprocessing and construction of a heterogeneous primitive library; S2, heterogeneous feature alignment and normalized representation; S3, construction of a dynamic routing decision space and generation of routing actions; S4, updating the reward network and routing strategy based on bi-layer primitive optimization; S5, updating the stable strategy based on reference strategy regularization; and S6, performing dynamic feature aggregation, task prediction, and controllable inference. This technical solution addresses the problems of fixed static network computation processes, simple dynamic network decision rules, difficulty in stably connecting heterogeneous feature modules, difficulty of manually adjusting manual reward functions to adapt to non-stationary routing optimization processes, and lack of runtime computational adjustment capabilities during the deployment phase. It constructs a dynamic backbone combination framework driven by data features, routing status, and deployment preferences. Unlike traditional methods that use the same network path for all input samples, this invention dynamically selects the appropriate feature extraction module, computation path, and exit position for the current sample based on changes in target scale, target density, background complexity, and current computational resource constraints in the remote sensing image. This reduces redundant computation in simple background areas and enhances the feature representation capability of complex target areas. The steps are described in detail below with reference to the accompanying drawings.

[0025] S1. Input preprocessing and heterogeneous primitive library construction: Receive high-resolution remote sensing images and perform standardized preprocessing. Construct a heterogeneous feature extraction primitive library containing convolution primitives, self-attention primitives, multi-scale primitives, lightweight primitives and hybrid structure primitives. Configure unified input and output interfaces and computational cost attributes for each primitive. This step is fundamental to the entire method. It is used to convert the original remote sensing image into a normalized tensor and to pre-build a candidate library containing various heterogeneous feature extraction primitives, providing selectable and composable computational units for subsequent dynamic routing.

[0026] S1.1 Tensorization Input of Multi-Scale High-Resolution Remote Sensing Images

[0027] like Figure 2As shown in node ①, the primary execution step in this embodiment is to receive and preprocess the raw visual payload data from the front-end Earth observation platform. This input data specifically refers to high-resolution visible light optical remote sensing images covering a wide geographic area. Given that real geographic environments typically exhibit significant spatial heterogeneity, and remote sensing images generally possess inherent characteristics such as extremely uneven information density distribution and vastly different target physical scales—for example, they may contain large port facilities, bridges, or land-water interfaces at the hundred-meter level, or simultaneously contain tiny vehicles, aircraft, or other weak targets occupying only a few pixels or even sub-pixel areas—it is essential to convert the original two-dimensional pixel matrix into a standardized mathematical tensor form before performing deep feature analysis. This ensures that subsequent heterogeneous feature extraction primitives can perform stable feature mapping and semantic modeling under a unified data interface. This module defines the input raw remote sensing image as a multidimensional data tensor X∈ H×W×3, where H and W represent the spatial resolution of the image in the height and width directions, respectively, and the symbol is... Representing the real number field, the constant 3 corresponds to the three physical color channels (red, green, and blue) in visible light imaging. In other words, each element of tensor X corresponds to the three-channel optical response value at a certain spatial location. This tensor representation is not only the unified input form for all subsequent feature extraction, structure mapping, and dynamic routing operations in this invention, but also provides a common mathematical carrier for data exchange between different heterogeneous operators. Through this unified tensor expression, the system can simultaneously incorporate local texture details, global spatial layout, and multi-scale target distribution from the same remote sensing image into the subsequent feature learning process, thereby driving multi-dimensional, multi-scale representation mapping and information deconstruction.

[0028] S1.2 Pre-construction of Heterogeneous Feature Extraction Primitive Library

[0029] like Figure 2 As shown in node ②, to overcome the rigidity and lack of adaptability of traditional homogeneous network computation graphs in complex remote sensing scenarios, this invention pre-instantiates and rigorously encapsulates a heterogeneous feature extraction primitive library at the system's bottom layer. This primitive library consists of multiple independent, parallelizable, and functionally complementary core computational families. Each primitive module has clearly defined input / output interfaces, computational power consumption attributes, and data flow specifications, thus providing sufficient structural freedom for subsequent dynamic backbone assembly and strategy control.

[0030] The first family consists of local texture and boundary extraction modules based on local receptive fields. These modules deploy depthwise separable convolutional blocks with different kernel sizes, residual feature aggregation units, and multi-scale convolution operators. Their core feature lies in prioritizing the modeling of high-frequency spatial geometric information in the image by relying on a local inductive bias mechanism. The underlying feature mapping process of this type of module can be abstractly represented as follows: in, Represents the local feature extraction mapping function. This represents a local feature representation generated by convolutional primitives. This type of mapping can focus on enhancing fine-grained information such as edges, textures, corners, and local topology, making it particularly suitable for finely depicting the physical contours of small and dense targets in remote sensing images and improving the discernibility of local structures in complex background clutter.

[0031] The second family consists of global semantic and long-range dependency modeling modules based on self-attention mechanisms. These modules integrate a sliding window attention operator supporting parallel processing, a global multi-head self-attention module, and a cross-dimensional label hybrid structure. Essentially, they overcome the receptive field limitations of local pixel grids, establishing global dependencies between features over a larger spatial span. The feature mapping process of this type of module can be represented as follows: ,in This represents a transformation operator that converts the original input or intermediate features into a serialized labeled representation. This represents a global semantic mapping function based on a self-attention mechanism. This represents the global semantic features output by the visual transformer class primitives. With this mapping mechanism, the system can calculate the correlation strength between pixels or regions on a large spatial scale, thereby more effectively modeling long-range semantic relationships and global context structures in complex urban road networks, large port areas, land-water boundaries, and multi-category target coupled distribution scenarios.

[0032] Therefore, the two types of heterogeneous primitives mentioned above are functionally complementary: the former excels at characterizing local boundaries and high-frequency details, while the latter excels at modeling global semantics and cross-regional dependencies. Together, they constitute the heterogeneous operator foundation of the underlying backbone composite architecture of this invention, providing ample structural candidate space for subsequent dynamic routing decisions.

[0033] In addition, the primitive library can also contain multi-scale primitives, lightweight primitives (such as MobileNet blocks), and hybrid structure primitives. Each primitive is configured with a uniform input / output interface and predefined computational cost attributes (such as FLOPs, number of parameters, or inference latency).

[0034] S2. Heterogeneous Feature Alignment and Standardized Representation: Through the feature topology alignment interface, the spatial scale, channel dimension and representation of the features output by different candidate primitives are transformed, converting convolutional spatial grid features and self-attention sequence features into a unified standardized feature representation; Bidirectional conversion of feature topology alignment interface: such as Figure 2As shown in node ③, a key challenge arises during the dynamic temporal concatenation of the two types of heterogeneous operators: the continuous spatial grid features output by the local feature extraction module and the discrete sequence features relied upon by the global semantic modeling module have inherent data structure differences. Without a unified, smooth, and reversible data alignment mechanism, spatial granularity mismatch, semantic mapping discontinuities, and gradient propagation discontinuities can easily occur during heterogeneous module cascading. To address this problem, this invention natively introduces a unified feature topology alignment interface at a critical point in the data flow chain. This interface is used to project the spatial grid features output by convolutional primitives forward into serialized marker features that can be processed by self-attention primitives, or to reconstruct the serialized marker features output by self-attention primitives backward into spatial grid features that can be processed by convolutional primitives.

[0035] Let the characteristic tensor of a continuous spatial grid in a certain intermediate layer be represented as: The corresponding discrete feature label sequence tensor is represented as ,in Indicates the number of valid tokens. This represents the hidden feature dimension of a single label. The alignment interface is responsible for performing bidirectional, reversible, and as lossless as possible structural transformations between spatial grid representations and sequence representations.

[0036] In the forward projection process from grid features to sequence features, the alignment interface achieves structural transformation through image patch embedding combined with positional encoding, which is mathematically expressed as follows:

[0037] in, This represents an operator that divides a two-dimensional feature map into several local blocks and maps them to sequence labels. Indicates embedding transformation, This represents the positional encoding term, used to preserve the relative topological relationships of the original spatial features in their physical locations. Through this process, a multi-channel two-dimensional feature map is transformed into a one-dimensional feature sequence, while its spatial layout information can still be explicitly preserved through positional encoding.

[0038] In the inverse reconstruction process from sequence features to grid features, the alignment interface restores the discrete sequence to two-dimensional spatial features through de-labeling mapping, which is mathematically expressed as:

[0039] in, This represents the inverse structure recovery operator, which can be implemented by pixel shuffling, channel rearrangement, deconvolution, or other structure recovery modules. The goal of this process is to reorganize high-order discrete sequence features into a spatial feature map with topological consistency, so that subsequent convolutional primitives can continue to perform local feature modeling on a regular grid representation.

[0040] The aforementioned bidirectional conversion mechanism fundamentally ensures the structural coherence of heterogeneous feature streams during forward propagation, as well as the smoothness and stability of gradient updates during backward propagation. Through this feature topology alignment interface, different heterogeneous primitives can achieve reversible, continuous, and low-loss data exchange under a unified protocol, providing reliable underlying communication support for subsequent dynamic backbone assembly.

[0041] Standardized feature output: such as Figure 2 As shown in node ④, after structural normalization and dimensional alignment are completed through the feature topology alignment interface, the outputs are further unified into standardized internal feature representations. This standardized output can be defined as... in, This represents the output characteristics of the original heterogeneous module. This represents the uniform feature alignment operator. This represents the standardized feature representation after alignment. The purpose of this process is to map features that may originally have different scales, different dimensional structures, and different semantic granularities to a unified feature space that can be consistently processed by the subsequent policy network, routing module, and task head.

[0042] Through this standardized feature output mechanism, the system effectively breaks down the underlying physical barriers in the process of assembling heterogeneous modules, significantly mitigating information loss caused by granularity mismatch, dimensional inconsistency, and semantic decay. More importantly, under the constraints of a unified interface protocol, various independent computing primitives can achieve true "plug-and-play" functionality, enabling the module-combining agents in the system's decision layer to freely explore a broader backbone combination space without introducing additional complex alignment overhead. Simultaneously, this standardized feature interface provides the necessary structural prerequisites for subsequent elastic deployment, runtime dynamic routing, and early exit mechanisms, ensuring that the entire invention is not only learnable and scalable during the training phase but also possesses stable, controllable, and reconfigurable engineering application capabilities during the actual deployment phase.

[0043] S3. Construct a dynamic routing decision space and generate routing actions: Construct a state representation based on current intermediate features, historical routing trajectories, network layer positions, candidate primitive states, consumed computing resources, and preference parameters; define candidate primitive selection, path switching, skipping layers, and early exit as action sets; the module-combined agent outputs the action probability distribution based on the state representation to generate a dynamic feature extraction path for the current input sample; This step models the dynamic feature extraction process as a Markov decision process, and the module-based intelligent agent selects the next action (primitive or early exit) based on the current state.

[0044] S3.1 State Space Design

[0045] like Figure 3 As shown, the decision agent of the dynamic feature extraction network is modeled as a Markov decision process, whose state space... It carries the network's current global context, historical trajectory, hierarchical position, and operational preference information, and is the basis for the agent to make optimal decisions.

[0046] Formalized, any time step The state is defined as: .

[0047] Global semantic feature representation: for the current intermediate feature tensor Perform global semantic pooling to generate low-dimensional but information-dense representation vectors. Its calculation formula is ,in For global average pooling, For normalized geometric feature encoding, It is a lightweight self-attention transformation.

[0048] Select a trajectory for the historical module and record the agent at time step. The selected module sequence. Its calculation formula is: in, For action, It is a trainable gating signal unit.

[0049] This is a deep-level location embedding used to provide network-level context-assisted decision-making. (For the current network depth...) Perform position encoding: .

[0050] It is a continuous preference label used to quantify the user's preference for detection accuracy. Computational cost Preferences: . Both the training and deployment phases can be flexibly adjusted to control the agent's decisions on the width and depth of feature extraction, achieving dynamic adjustment to balance accuracy and efficiency, and supporting seamless switching of a single training model under different computing power constraints.

[0051] The state-space fusion mechanism is the core component of the dynamic heterogeneous backbone network design in this invention. Its goal is to systematically integrate heterogeneous information from four dimensions to form a high-dimensional state vector that fully reflects the current network state, environmental constraints, and user preferences, providing a comprehensive and reliable decision-making basis for the reinforcement learning agent. In the initial fusion stage, the four types of state information are nonlinearly mapped and concatenated using a multilayer perceptron (MLP):

[0052] This process not only achieves information dimension unification but also enhances the discriminative power of the state vector through activation functions, enabling the agent to perceive the environmental state simultaneously at both macro and micro scales. The feature-level fusion stage ensures that the interaction information between historical trajectories, hierarchical location information, and current features can be fully encoded, providing rich initial representations for subsequent attention mechanisms.

[0053] In the fused vector Furthermore, a self-attention mechanism is applied, followed by layer normalization:

[0054] Attention enhancement can capture dependencies across different state dimensions, such as the importance of historical module selection to current features, or the guiding role of preference vectors in hierarchical decisions. This mechanism ensures that the agent can consider both long-term memory and local contextual features when making dynamic routing choices, enhancing the global consistency and robustness of state representations.

[0055] Ultimately, the state vector must be adaptable to different types of feature extraction modules: ① Convolutional modules: emphasizing spatial resolution information and local boundary features, enabling the agent to precisely preserve microstructures when selecting convolutional blocks. ② Self-attention modules: emphasizing global dependencies across space and channels, enabling the agent to fully model macroscopic semantic relationships when selecting attention blocks. Through this perceptual mechanism, the fused state vector can provide effective decision guidance when facing heterogeneous modules, achieving a balance between high accuracy and high efficiency.

[0056] Action space design and dynamic routing strategy: In this invention, the action space... It is the core of reinforcement learning agents for dynamic architecture generation, and its design directly determines the selection, combination, and execution accuracy of heterogeneous modules and early termination mechanisms. The action space is precisely defined as a set of discrete decision instructions, each action... Corresponding to specific module selection or control operations, ensure that at any time step Internally, the agent can rely on the state vector Make the optimal route selection.

[0057] The motion space mainly includes two types: 1) Module selection action: The agent selects modules from a pre-built heterogeneous module toolbox. Selecting the next feature extraction primitives:

[0058] in Represents a convolutional module, attention module, or hybrid module that performs a specific function; This indicates an early exit action, used to terminate the current feature extraction process. The selection decision for each module depends not only on the current state. It also incorporates historical selection trajectories. and hierarchical embedding Perform global context awareness to ensure coordination between macroscopic and microscopic features.

[0059] 2) Early Exit Action: The agent determines that the current accumulated features already meet the user's preferences. When specified precision requirements are met, the following can be executed. action:

[0060] This action triggers a truncation of the feature extraction process, avoiding redundant computation while ensuring the semantic integrity of the accumulated features is directly passed to the task detection head. The early exit mechanism is a key element in achieving elastic control of computing resources and controllable inference at runtime.

[0061] Action selection strategy: at each time step Internally, the agent determines the current state vector. Output policy distribution:

[0062] in The action probability distribution predicted by the policy network; A multilayer perceptron is used to parameterize the policy function. The action set follows a multivariate distribution Sampling.

[0063] The action selection process follows a probabilistic sampling mechanism, enabling the agent to strike a balance between exploring efficient paths and maintaining high-precision features.

[0064] During the training phase, random sampling ensures that the agent can explore different combinations of paths, avoiding premature convergence to a local optimum. During the deployment phase, a greedy strategy can be used to select the optimal action to ensure task efficiency.

[0065] Dynamic routing and module composition: Actions for each selected module The corresponding operation will be applied to the current input features. :

[0066] in This represents the forward feature transformation function of the module. For different types of heterogeneous modules: convolutional modules emphasize local spatial feature extraction while preserving microscopic boundary information; attention modules capture long-distance dependencies and macroscopic semantics through a global self-attention mechanism; hybrid modules combine convolutional and attention features to balance local and global representation capabilities.

[0067] Modules communicate via a feature alignment interface Ensure consistent spatial resolution and channel dimensions for heterogeneous outputs:

[0068] Through interface mapping, convolutional outputs and attention outputs can be smoothly concatenated or chained, ensuring the integrity and differentiability of features throughout the combination path. When the proxy executes... During the action, the current accumulated features The data is routed to the detection head, enabling instance-aware dynamic feature extraction.

[0069] This design enables a single model to adaptively select the computation path based on its state and preferences during runtime, balancing computational efficiency and detection accuracy.

[0070] Motion space design principles: The design of motion space should follow these principles: 1) Comprehensiveness: Covers all possible feature extraction modules and early termination actions to ensure the integrity of the search space; 2) Composability: Allows different types of modules to be dynamically assembled according to strategies, giving full play to the advantages of heterogeneous features; 3) Flexibility: Combining continuous preference vectors Adjusting module depth and early exit probability achieves a dynamic trade-off between high precision and high efficiency; 4) Controllability: During the training phase, internal reward signals are used. Guided strategy updates allow for direct control of action distribution during the deployment phase through preference adjustments.

[0071] S4. Update the reward network and routing strategy based on two-layer meta-optimization: In the outer layer optimization process, the meta-reward network generates an intrinsic reward signal based on the state-action trajectory, detection loss, computational cost, and preference parameters, and updates the meta-reward network parameters based on the comprehensive evaluation target of the validation set; In the inner layer optimization process, the module-combined agent updates the routing strategy parameters based on the intrinsic reward signal. This invention employs a two-layer cyclic strategy optimization mechanism, which achieves global optimization of the state-action space and long-term performance guidance through the collaborative operation of the inner reinforcement learning policy network and the outer meta-network.

[0072] Inner layer optimization: The inner reinforcement learning agent uses the fused high-dimensional state vector As input, based on the current policy distribution Execute action And based on the intrinsic reward generated by the outer meta-network For strategy parameters Iterative updates are performed. The design goal of the inner RL layer is to enable the policy network to effectively explore heterogeneous module sequences in the discrete combinatorial space, while ensuring that the accumulated features meet the task accuracy requirements and take into account computational efficiency.

[0073] Trajectory Sampling: In each training iteration, the agent generates a complete state-action trajectory by interacting with the environment.

[0074] in, The fused state vector contains global semantic features, historical module selection, hierarchical embedding, and continuous preference labeling. Select or exit the module early; Generated by the outer meta-network, it measures the combined contribution of the current action to accuracy and computational efficiency. Trajectory sampling ensures that the agent can fully explore a broad dynamic module combination space, while providing complete decision information for policy gradient calculation.

[0075] Inner objective function: To prevent policy collapse and reward cheating, the inner RL objective function introduces KL divergence regularization based on the reference policy, in addition to the traditional cumulative reward:

[0076] in: For the current policy network; The reference strategy for the previous round (e.g., behavior cloning or the optimal strategy for the previous round). The regularization weights are used to ensure that the policy maintains similarity to the distribution of the reference policy while pursuing the maximization of intrinsic rewards, thus suppressing early exit or degenerate decisions caused by local extrema of short-term rewards.

[0077] Policy gradient update: agent parameters Update via policy gradient method:

[0078] in: To accumulate rewards for discounts, This serves as a discount factor, controlling the memory weight of the agent in long-term decision-making.

[0079] Dynamic routing execution: In the inner loop, each selected module... The forward operation is:

[0080] When performing an early exit action:

[0081] Through the above mechanism, the inner reinforcement learning agent can achieve instance-aware dynamic module concatenation and resource-adaptive reasoning.

[0082] Outer layer optimization: The core of outer layer meta-optimization lies in using meta-networks. Generate intrinsic reward signals It directly optimizes the global performance of the inner policy on the validation set, thereby overcoming the limitations of the static reward function and guiding the policy to converge to a high-precision and high-efficiency global optimal solution during training.

[0083] Meta-network modeling: The meta-network adopts a gated network approach.

[0084] The input is the inner proxy trajectory. The output is a continuous intrinsic reward signal, used to measure the contribution of state-action combinations to global performance.

[0085] Meta-optimization objective: The meta-objective is to maximize the joint optimality of accuracy and computational efficiency on the validation set.

[0086] Meta-gradient Update: The meta-network parameters are updated by calculating the meta-gradient using the first-order Taylor approximation.

[0087] This process ensures that the intrinsic reward signal can dynamically adjust the inner policy, enabling it to optimize stably even in non-stationary reward environments.

[0088] Inner and outer layer interaction: Inner layer strategy based on state Generate Actions The outer layer meta-network is based on the trajectory Output The inner layer is updated; the inner layer strategy is iteratively updated under KL regularization constraints, while continuously exploring and utilizing; the validation set evaluation results are fed back to the meta-network to optimize the intrinsic reward for the next round.

[0089] S5. Stable policy update based on reference policy regularization: Introduce Kullback-Leibler divergence constraints between the current routing policy and the reference routing policy in the inner policy optimization objective to limit the difference in the output action probability distribution of the two under the same state; To reduce policy fluctuations caused by continuous changes in the meta-reward signal, this step explicitly incorporates the current policy into the inner-layer optimization objective. Reference Strategy The KL divergence term between them (already reflected in the inner objective function of step S4). The reference strategy can be the parameters of the previous round, the historical average strategy parameters, or the exponential moving average strategy parameters.

[0090] Specifically, the complete inner objective function is:

[0091] This regularization constraint ensures that while pursuing high rewards, the strategy does not deviate excessively from a stable reference strategy, thereby preventing premature degradation into a single shallow path or the generation of violent oscillations.

[0092] S6. Perform dynamic feature aggregation, task prediction and controllable inference: Perform scale unification, channel mapping, local detail enhancement and global context aggregation on the features output by different dynamic paths to obtain unified task features and input them into the adaptive detection head to complete task prediction; during the deployment phase, adjust the computation path according to preference parameters and budget constraints to achieve controllable inference at different accuracy and efficiency working points.

[0093] like Figure 4 As shown, an adaptive feature gateway is first introduced to perform unified normalization mapping on dynamic features generated at any exit level. This gateway is essentially a feature interface coupled with a dynamic channel projection operator and a spatial resampling operator, aiming to output arbitrary paths without destroying the original task information. Mapped to a predefined standard feature subspace

[0094] in , , Standard channel size and standard spatial resolution are defined uniformly for the system. Specifically, this can be expressed as...

[0095] in, and These represent the channel projection parameters that are adapted to the current input feature attributes. This represents a convolution or linear projection operation. Represents a non-linear activation function. This represents an operator that resamples features to a standard spatial scale.

[0096] Dynamic feature aggregation: further processing of standardized features Dynamic feature aggregation is performed to achieve on-demand enhancement and task-oriented refinement of target information. Unlike traditional fixed feature pyramid networks, this invention does not pre-define several static scale branches for mechanical fusion. Instead, it adaptively allocates aggregation weights for local and global enhancement based on the actual response distribution, saliency density, and contextual dependency strength of the feature map corresponding to the current sample. Let the standardized features be... The system first generates a spatial response map and channel weight vectors using a lightweight response evaluation network:

[0097] in, This indicates the potential contribution intensity of each spatial location to the mission. This represents the semantic activation weights of different channels.

[0098] Based on this, the system constructs local detail enhancement branches and global context enhancement branches respectively.

[0099] Local branches can be written as

[0100] in It can be composed of shallow convolution, deformable convolution or edge-sensitive convolution, and its main task is to further highlight the edges of small targets, the contours of slender targets and the subtle separation relationships between dense instances.

[0101] Global branches are represented as

[0102] in By employing large receptive field convolution, global attention, or cross-regional relationship modeling operators, the semantic coupling between distant regions is enhanced, making it particularly suitable for targets with strong overall topological attributes, such as bridges, roads, waterways, airport runways, and large ships. Subsequently, the system calculates the adaptive fusion coefficients of the two branches based on the content complexity and feature response state of the current sample. And obtain the final aggregated features. .

[0103] Adaptive detection head: In obtaining aggregated features The system then feeds the data into an adaptive detection head to complete the final category determination, location regression, and optional pixel-level dense prediction tasks. To adapt to the characteristics of remote sensing images, such as arbitrary target orientation, large aspect ratio variations, and significant scale differences between categories, this embodiment preferably uses a structurally decoupled lightweight detection head, which models the classification branch, geometric regression branch, and segmentation branch in parallel.

[0104] The classification branch outputs the class probability at each candidate position. , This represents the confidence distribution of the target belonging to each predefined category.

[0105] Regression branch output target box parameters For a regular rectangle, it can be represented as For directional targets, which are more common in high-resolution remote sensing, this invention further supports rotated bounding box modeling, i.e. , The target orientation angle is used to characterize the arbitrary orientation features of ships, aircraft, vehicles, port facilities, etc. in top-down imaging.

[0106] If the task requires pixel-level dense interpretation, the segmentation branch further outputs a semantic mask or an instance mask.

[0107] From a system representation perspective, the execution process of a traditional static network during the deployment phase can typically be written as follows:

[0108] in For the input image, To fix the model parameters, This is the output result. This form implies that once training is complete, the model's computation graph, path lengths, calling modules, and resource consumption patterns are completely fixed and cannot be adjusted in real-time according to the external environment. However, in this invention, the inference process during the deployment phase is reconstructed into a conditional inference process constrained by runtime control variables, which can be expressed as follows: ,in, This represents the network parameters of the dynamic decision-making strategy after freezing. This represents the parameters of the frozen meta-control network. This indicates parameters related to backend aggregation and detection headers. This indicates that the control vector is injected in real time by the external system during the deployment phase. In other words, what is fixed during deployment in this invention is the "parameters" rather than the "computation path"; the same set of parameters can correspond to different depths, complexities, and module combinations of dynamic backbone execution modes under the action of different control vectors, thereby giving the model flexibility in the real operating environment.

[0109] Runtime preference control and conditional reasoning mechanisms, such as Figure 5 As shown in this embodiment, in order to enable the model's reasoning behavior to be continuously adjusted, the system encodes the external task objectives and resource preferences into a low-dimensional control vector and embeds it into the state space of the decision agent.

[0110] Preferably, the runtime control vector can be defined as follows:

[0111] in, This indicates the strength of the preference for recognition accuracy. This indicates the degree of constraint on inference latency. This indicates a preference for energy consumption control. This indicates the sensitivity to video memory or on-chip cache usage. To facilitate unified cross-platform access, the above components are typically normalized to meet certain requirements. In a simplified implementation, only the two main variables, "accuracy preference" and "efficiency preference," can be retained. ,in The larger the value, the more the system tends to call deeper, more semantically complex feature extraction modules; The larger the value, the more the system tends to compress path length, reduce module calls, and exit as early as possible. Regardless of whether a multivariate or simplified bivariate form is used, the essence is to transform external demands into conditional signals that can be perceived and utilized by the neural network.

[0112] Control Vector At runtime, the action distribution, together with the current feature state, constitutes the input to the policy network; therefore, the original action distribution... Expanded to conditional form ,in For the first The state representation at each decision moment. The actions to be executed include continuing to call a specific convolutional primitive, an attention primitive, a hybrid primitive, or directly triggering an exit action. Therefore, the policy network no longer determines subsequent paths solely based on "how complex the current image looks," but further considers the runtime objective of "whether the current system prioritizes accuracy or efficiency," thus achieving truly controllable inference. For the same remote sensing image, in Depending on the value, the system may produce drastically different dynamic execution trajectories. ,in This indicates the actual execution depth under control vector constraints. This characteristic means that the model itself does not need to maintain multiple structural versions; a single model can continuously switch between various inference modes by adjusting external control parameters.

[0113] Budget-Constrained Elastic Path Scheduling: To further align deployment behavior with real-world hardware constraints, an explicit path scheduling mechanism oriented towards computing budgets is introduced, building upon preference control. In actual deployments, the available latency, energy consumption, and storage budgets of ground servers, vehicle-mounted computing units, airborne edge platforms, and spaceborne processors often vary significantly. Therefore, this invention does not rely solely on abstract preferences but allows the system to receive specific budget parameters for the current platform and continuously perceive and update them during the decision-making process.

[0114] Assume the system's current acceptable budget constraint is

[0115] in This represents the maximum time delay budget for a single image or a single sub-region. This indicates the allowable energy consumption budget. Indicates the available storage budget. Suppose a certain action... The corresponding module call cost is

[0116] Then proceed along the current path to the [number]th [path]. At step, the remaining budget of the system can be denoted as... .

[0117] Therefore, the dynamic decision-making process is a condition-controlled process constrained by the three factors of "content, preference, and budget," that is... Under this expression, the policy network not only knows the semantic complexity of the current image and whether the current system is biased towards high precision or high efficiency, but also knows how many resources are still available for consumption. When the budget is relatively abundant, the policy can prioritize global attention modules, hybrid aggregation modules, or deeper feature extraction links with stronger expressive power but higher costs; while when the budget is tight, the policy is more likely to tilt towards low-cost modules or exit actions early, thereby obtaining the best possible task output under given constraints. In a preferred implementation, the system can also impose a budget feasible region constraint on the action space. Let the... The set of actions that can still be executed during a step is Then action sampling is restricted to This is done internally. In this way, the system naturally possesses "executability awareness" during runtime, avoiding invalid paths that, while semantically warranting further exploration, are unmanageable by the hardware. Unlike traditional static models that "run the entire process first and then accept results regarding timeouts or excessive energy consumption," this invention continuously performs resource scheduling and path correction during the forward inference process, thus directly embedding budget constraints into the model's execution logic.

[0118] Hierarchical controllable inference for different scenario complexities, specifically as follows: Figure 5 As shown, remote sensing images are often extremely large in size, and the complexity of different regions within the same image varies significantly. Therefore, the flexible deployment of this invention is not limited to unified inference across the entire image, but also further supports hierarchical control for different regions. In actual operations, an image may simultaneously contain various heterogeneous regions such as sea surfaces, mountains, urban areas, ports, roads, and farmland. Some regions contain almost no high-value targets, while others are densely populated with a large number of small targets or complex topological targets. If the deepest path is uniformly used for interpretation across all regions, significant redundant computation will occur; if the shallowest path is uniformly used, the recognition performance of complex regions will be insufficient. To address this, this invention supports a low-cost pre-analysis module to roughly estimate the scene complexity before full-image-level inference, and then divides the original image into several local regions and assigns different control parameters to each.

[0119] Suppose the input image is divided into The system can then generate local control vectors for each region based on its saliency, texture density, candidate target distribution, and current budget state. ,in For runtime control mapping functions, This indicates the current task priority or category of interest. Subsequently, each sub-region is defined by its own... By performing conditional inference using the same frozen model, different dynamic execution depths can be obtained. .

[0120] For areas with simple backgrounds, uniform textures, and few prominent targets. The system will prioritize efficiency-oriented control parameters, thereby triggering shallower computations and earlier exits; while for areas suspected of containing dense targets, complex transportation networks, large naval formations, or high-value facilities... This outputs a control vector that emphasizes precision, guiding the strategy to choose a path with deeper and stronger semantic modeling capabilities. Through this mechanism, the present invention can not only achieve controllable inference at the model level, but also achieve differentiated resource allocation in different regions within the image, allowing the computational budget to be prioritized for truly complex and important regions, significantly improving overall throughput efficiency and effective recognition rate.

[0121] Example 2: A reinforcement learning-based backbone module system for remote sensing images

[0122] This invention provides a reinforcement learning ensemble backbone module system for remote sensing images, comprising: The heterogeneous primitive library construction module is used to execute step S1, construct a heterogeneous feature extraction primitive library containing at least two types of feature extraction primitives, and configure a unified input / output interface and computational cost attribute for each primitive; Heterogeneous Feature Alignment Module: Used to perform step S2, which standardizes the spatial, channel, and representational forms of heterogeneous features output by different primitives; Dynamic routing decision module: used to execute step S3, including module combination agent, used to construct state representation and action space, and generate dynamic feature extraction path; Two-layer meta-optimization module: used to execute step S4, including a meta-reward network for outer-layer meta-reward learning and inner-layer routing strategy optimization; Policy stabilization module: Used to execute step S5, introduce reference policy KL divergence regularization, and stabilize the policy update process; Controllable Inference and Flexible Deployment Module: Used to execute step S6, receive deployment preference parameters during the deployment phase, and control the trained module combination agent to dynamically select feature extraction paths for the input image.

[0123] The dynamic feature aggregation and adaptive detection head module is used to execute step S6, which aggregates the dynamic path output features and completes target detection, orientation angle prediction or instance segmentation tasks through the adaptive detection head.

[0124] Finally, it should be noted that although the present invention has been described with reference to specific embodiments, those skilled in the art should recognize that the above embodiments are only used to illustrate the present invention and are not intended to limit the present invention. Various equivalent changes or substitutions can be made without departing from the concept of the present invention. Therefore, any changes or modifications to the above embodiments within the essential spirit of the present invention will fall within the scope of the claims of the present invention.

Claims

1. A reinforcement learning-based backbone module combination method for remote sensing images, characterized in that, Includes the following steps: S1. Input preprocessing and heterogeneous primitive library construction: Receive high-resolution remote sensing images and perform standardized preprocessing. Construct a heterogeneous feature extraction primitive library containing convolution primitives, self-attention primitives, multi-scale primitives, lightweight primitives and hybrid structure primitives. Configure unified input and output interfaces and computational cost attributes for each primitive. S2. Heterogeneous Feature Alignment and Standardized Representation: Through the feature topology alignment interface, the spatial scale, channel dimension and representation of the features output by different candidate primitives are transformed, converting convolutional spatial grid features and self-attention sequence features into a unified standardized feature representation; S3. Construct a dynamic routing decision space and generate routing actions: Construct a state representation based on current intermediate features, historical routing trajectories, network layer positions, candidate primitive states, consumed computing resources, and preference parameters; define candidate primitive selection, path switching, skipping layers, and early exit as action sets. The modular intelligent agent outputs the action probability distribution based on the state representation, and generates a dynamic feature extraction path for the current input sample; S4. Update the reward network and routing strategy based on two-layer meta-optimization: In the outer layer optimization process, the meta-reward network generates an intrinsic reward signal based on the state-action trajectory, detection loss, computational cost, and preference parameters, and updates the meta-reward network parameters based on the comprehensive evaluation target of the validation set; In the inner layer optimization process, the module-combined agent updates the routing strategy parameters according to the intrinsic reward signal. S5. Stable policy update based on reference policy regularization: Introduce Kullback-Leibler divergence constraints between the current routing policy and the reference routing policy in the inner policy optimization objective to limit the difference in the output action probability distribution of the two under the same state; S6. Perform dynamic feature aggregation, task prediction and controllable inference: Perform scale unification, channel mapping, local detail enhancement and global context aggregation on the features output by different dynamic paths to obtain unified task features and input them into the adaptive detection head to complete task prediction; During the deployment phase, the computation path is adjusted based on preference parameters and budget constraints to achieve controllable inference at different accuracy and efficiency working points.

2. The reinforcement learning-based backbone module method for remote sensing images according to claim 1, characterized in that, The feature topology alignment interface in S2 is used to perform forward projection of the spatial grid features output by the convolutional primitive into serialized label features that can be processed by the self-attention primitive, or to reverse reconstruction of the serialized label features output by the self-attention primitive into spatial grid features that can be processed by the convolutional primitive.

3. The reinforcement learning-based backbone module method for remote sensing images according to claim 1, characterized in that, The state representation in S3 is generated by concatenating global semantic feature representation, historical module selection trajectory, deep hierarchical position embedding, and continuous preference label, followed by nonlinear mapping and self-attention enhancement; the continuous preference label is used to quantify the degree of preference for detection accuracy and computational cost, with a value range of [0,1] and the sum of the two is 1.

4. The reinforcement learning-based backbone module method for remote sensing images according to claim 1, characterized in that, In S4, the objective of the outer meta-optimization is to maximize the joint optimality of detection accuracy and computational efficiency on the validation set; the inner reinforcement learning uses the policy gradient method to update the routing policy parameters, and the objective function is to introduce KL divergence regularization based on the reference policy on the basis of traditional cumulative reward.

5. The reinforcement learning-based backbone module method for remote sensing images according to claim 1, characterized in that, The reference routing strategy in S5 is determined by the strategy parameters of the previous round, the historical average strategy parameters, the exponential moving average strategy parameters, or the strategy parameters saved in the stable phase.

6. The reinforcement learning-based backbone module method for remote sensing images according to claim 1, characterized in that, The dynamic feature aggregation process in S6 includes: The adaptive feature gateway maps the dynamic features output from any path to a preset standard feature subspace. Based on the feature response distribution, a spatial response map and channel weight vector are generated, and local detail enhancement branches and global context enhancement branches are constructed respectively. The fusion coefficients of the two branches are adaptively calculated based on the complexity of the sample content to obtain the final aggregated features.

7. The reinforcement learning-based backbone module method for remote sensing images according to claim 1, characterized in that, The controllable reasoning process in S6 includes: The runtime control vector is input into the policy network. The control vector includes the preference for recognition accuracy, the degree of constraint on inference latency, the preference for energy consumption control, and the sensitivity to the occupation of video memory or on-chip cache. During the reasoning process, the remaining computational budget is continuously perceived, and action sampling is restricted to the budget feasible region; It supports dividing remote sensing images into multiple local regions and assigning different control parameters according to the complexity of each region to achieve hierarchical and controllable inference.

8. A reinforcement learning ensemble backbone module system for remote sensing images, characterized in that, include: The heterogeneous primitive library construction module is used to build a heterogeneous feature extraction primitive library containing at least two types of feature extraction primitives, and to configure a unified input / output interface and computational cost attribute for each primitive. Dynamic routing decision module: contains a module combination agent, used to construct state representation and action space, and generate dynamic feature extraction paths; Two-layer meta-optimization module: includes a meta-reward network, used for outer-layer meta-reward learning and inner-layer routing strategy optimization; Policy stabilization module: Used to introduce reference policy KL divergence regularization to stabilize the policy update process; Controllable Inference and Flexible Deployment Module: Used to receive deployment preference parameters during the deployment phase and control the trained module combination agent to dynamically select feature extraction paths for the input image.

9. A reinforcement learning ensemble backbone module system for remote sensing images according to claim 8, characterized in that, Also includes: Heterogeneous Feature Alignment Module: Used to standardize the spatial, channel, and representational forms of heterogeneous features output by different primitives; The dynamic feature aggregation and adaptive detection head module is used to aggregate the output features of dynamic paths and complete target detection, orientation angle prediction or instance segmentation tasks through the adaptive detection head.

10. A reinforcement learning ensemble backbone module system for remote sensing images according to claim 9, characterized in that, The adaptive detection head in the dynamic feature aggregation and task prediction module adopts a structural decoupling design, which includes parallel classification branches, directional regression branches, and optional segmentation branches, which output class probabilities, rotated bounding box parameters, and pixel-level masks, respectively.

Images