A small target detection method and device based on dynamic state space and flow matching

Abstract

This scheme discloses a small target detection method and apparatus based on dynamic state space and flow matching. The method includes: extracting multi-scale feature maps from the original image; inputting the multi-scale feature maps into a two-dimensional selective scanning module for depth processing to obtain an enhanced feature map that integrates dynamic spatial attention and multi-scale perturbation information; introducing a spatial dimension local transfer mechanism into the enhanced feature map and fusing and updating the state vector of the enhanced feature map to obtain a spatial-temporal enhanced feature map; obtaining the target feature map through a feature space flow matching mechanism based on Lie index mapping; and inputting the target feature map into a detection head to obtain the target object's category, bounding box, and confidence score. This scheme supplements fine-grained feature information while preserving the original feature structure, retaining the ability to model long sequences, avoiding redundant computation of self-attention and the local receptive field limitation of traditional convolution, and realizing on-demand allocation of computational resources.

Description

Technical Field

[0001] This solution relates to the field of computer vision technology, and in particular to a method and apparatus for small target detection based on dynamic state space and flow matching. Background Technology

[0002] In space missions such as on-orbit servicing and space situational awareness, accurate detection of long-distance spacecraft and their key components is a prerequisite for achieving autonomous operation and intelligent decision-making. During long-distance detection, targets typically appear as "small targets" with a very low pixel ratio on imaging sensors. Their characteristic signals are weak, detailed information is sparse, and they are easily drowned out by noise from complex space backgrounds (such as Earth's albedo and the starry sky).

[0003] Existing methods suffer from three major bottlenecks in small object detection:

[0004] 1. The contradiction between efficiency and detail: The convolution operation of the YOLO series has strong locality but weak global correlation and insufficient modeling of the contextual dependency of small targets; Transformer-based methods (such as DETR) have strong global modeling capabilities, but their high complexity leads to poor real-time performance.

[0005] 2. Insufficient dynamic adaptability: Whether it is YOLO's fixed anchor points, Deformable DETR's predefined sampling points, or Mamba YOLO's fixed scanning direction, it is difficult to dynamically adjust the modeling granularity according to the density of small targets and the complexity of the background, resulting in "wasted computation in simple scenes and insufficient accuracy in complex scenes".

[0006] 3. Poor feature robustness: Small target features are easily affected by noise and occlusion. The activation function and channel weights of existing models are fixed and cannot adaptively enhance the expression of key feature channels.

[0007] Therefore, there is an urgent need for a small target detection technology that can achieve high efficiency, high accuracy, and high robustness under conditions of limited computing resources, in order to overcome the shortcomings of the existing technologies mentioned above. Summary of the Invention

[0008] This solution aims to at least address the technical problems existing in the prior art. To this end, the first aspect of this invention proposes a small target detection method based on dynamic state space and flow matching, the method comprising:

[0009] The original image is processed by a feature extraction module to extract multi-scale feature maps containing semantic and spatial information at different levels;

[0010] The multi-scale feature map is input into the two-dimensional selective scanning module for depth processing to enhance the key features of small targets and suppress noise, resulting in an enhanced feature map that integrates dynamic spatial attention and multi-scale perturbation information.

[0011] A spatial dimension local transfer mechanism is introduced into the enhanced feature map, and the state vector of the enhanced feature map is fused and updated to obtain a spatial-temporal enhanced feature map;

[0012] By using a feature space flow matching mechanism based on Lie index mapping, the sparse and scattered feature distributions in the spatial-temporal enhanced feature map are mapped and converged into a high-dimensional manifold region to obtain the target feature map;

[0013] The target feature map is input into the detection head, and the target feature map is processed by the classification branch and regression branch of the detection head to obtain the category, bounding box and confidence of the target object in the original image.

[0014] Optionally, the step of inputting the multi-scale feature map into a two-dimensional selective scanning module for depth processing to enhance the features of key regions of small targets and suppress noise includes:

[0015] The multi-scale feature map is projected through the first convolutional layer and then nonlinearly transformed using the first activation function to obtain the projected features.

[0016] The projected features are fed in parallel into the main feature processing branch and the multi-scale feature perturbation fusion branch for processing, to obtain the main features and multi-scale perturbation features respectively.

[0017] The core features and the multi-scale perturbation features are fused to obtain the fused features;

[0018] The fused features are input into the dynamic feature adjustment branch for processing to generate dynamically adjusted features;

[0019] The dynamically adjusted features are multiplied by the attention weight factor of the dynamically adjusted feature branch to obtain weighted features. The weighted features are then fed into the selective scanning module, where sequence modeling is performed using a parameterized state transition function. The hidden states are dynamically updated to capture long-distance dependencies between features, resulting in an enhanced feature map that integrates dynamic spatial attention and multi-scale perturbation information.

[0020] Optionally, the step of feeding the projected features into the main feature processing branch and the multi-scale feature perturbation fusion branch in parallel for processing, to obtain the main features and multi-scale perturbation features respectively, includes:

[0021] The projected features are sent to the main feature processing branch for processing to obtain the main features.

[0022] The projected features are fed in parallel into the first heterogeneous feature transformation path, the second heterogeneous feature transformation path, and the third heterogeneous feature transformation path of the multi-scale feature perturbation fusion branch;

[0023] The first feature is obtained by linearly transforming the projected feature through a 1×1 convolutional layer of the first heterogeneous feature transformation path;

[0024] The projected features are downsampled by a 3×3 depthwise convolution of the second heterogeneous feature transformation path, and after intermediate convolution processing, they are upsampled by bilinear interpolation to restore them to their original size, thus obtaining the second feature.

[0025] The number of channels of the projected feature is compressed to C' through a 1×1 convolution of the third heterogeneous feature transformation path, then spatial perturbation is introduced through a 3×3 depthwise convolution, and then the number of channels is restored to C through a 1×1 convolution to obtain the third feature; where C is the original number of channels of the projected feature pair, and C' is the number of channels after compression.

[0026] The first feature, the second feature, and the third feature are weighted and fused using a set of learnable weights to obtain multi-scale perturbation features.

[0027] Optionally, the fused features are input into a dynamic feature adjustment branch for processing to generate dynamically adjusted features, including:

[0028] The fused features are input into a 1×1 convolutional layer of the dynamic feature adjustment branch to compress the number of channels of the fused features, thus obtaining compressed features.

[0029] The compressed features are input into the activation function and subjected to nonlinear transformation to obtain the activation features;

[0030] The activation features are input into a 1×1 convolutional layer to restore the number of channels of the activation features, thus obtaining the restored features;

[0031] The recovered features are processed by the Sigmoid function to generate a spatial attention weight map with values ​​ranging from [0,1].

[0032] The spatial attention weight map is multiplied element-wise with the projected features to obtain dynamically adjusted features.

[0033] Optionally, the step of introducing a spatial dimension local transfer mechanism into the enhanced feature map and fusing and updating the state vector of the enhanced feature map to obtain a spatial-temporal enhanced feature map includes:

[0034] The spatial dimension local transfer module receives the previous state vector and the current input features of the enhanced feature map;

[0035] The state after the temporal evolution of the state vector at the previous moment is calculated based on the state transition function to obtain the temporal evolution state;

[0036] The spatial neighborhood information of the state vector at the previous time step is aggregated using the spatial transition operator, and the spatial transition spectrum radius is used as a constraint index in the state update process to obtain the spatially aggregated state.

[0037] The temporal evolution state and the spatial aggregation state are weighted and fused using a learnable spatial-temporal fusion coefficient to generate the current state vector and obtain the spatial-temporal enhanced feature map.

[0038] Optionally, the step of mapping and converging the sparse and scattered feature distributions in the spatial-temporal enhanced feature map into a high-dimensional manifold region through a feature space flow matching mechanism based on Lie index mapping to obtain the target feature map includes:

[0039] Obtain the original distribution of the single-frame small target feature set in the spatial-temporal enhanced feature map, and select a special Euclidean group as a Lie group. Map the original distribution to the compact discriminative distribution of large target features through Lie group transformation to obtain the transformed distribution.

[0040] By utilizing Lie group operations, the cross-scale features of the transformed distribution are collaboratively optimized to complete the fracture features of small targets and obtain optimized distribution features.

[0041] A Lie algebra regularization term is applied to the optimized distribution feature, which is obtained based on the gradient on the Lie group manifold; by minimizing the regularization term, the jumps of the optimized distribution feature on the high-dimensional manifold are reduced, and the target feature map is obtained.

[0042] Optionally, the step of utilizing Lie group actions to achieve collaborative optimization of the cross-scale features of the transformed distribution, completing the fracture features of small targets, and obtaining optimized distribution features includes:

[0043] The transformation from high-scale features to low-scale features is defined as a Lie algebra mapping;

[0044] Using the Lie algebra mapping, the distances of features at different scales on the Lie group manifold are calculated to obtain the Lie group distance;

[0045] Calculate the Lie group entropy of the feature distribution at each scale, and determine the fusion weight of each Lie group entropy based on the principle that the smaller the Lie group entropy, the higher the weight; according to the fusion weight, the features at each scale are weighted and fused, and in the process of weighted fusion, the consistency of features at different scales is constrained by minimizing the Lie group distance during the fusion process to obtain the optimized distribution features.

[0046] A second aspect of the present invention provides a small target detection device based on dynamic state space and feature flow matching, the device comprising:

[0047] The feature extraction module is used to extract multi-scale feature maps containing semantic and spatial information at different levels from the original image.

[0048] The depth processing module is used to input the multi-scale feature map into the two-dimensional selective scanning module for depth processing, enhance the key features of small targets and suppress noise, and obtain an enhanced feature map that integrates dynamic spatial attention and multi-scale perturbation information.

[0049] The fusion update module is used to introduce a spatial dimension local transfer mechanism into the enhanced feature map and to fuse and update the state vector of the enhanced feature map to obtain a spatial-temporal enhanced feature map.

[0050] The aggregation module is used to map and aggregate the sparse and scattered feature distributions in the spatial-temporal enhanced feature map into a high-dimensional manifold region through a feature space flow matching mechanism based on Lie index mapping, thereby obtaining the target feature map;

[0051] The detection module is used to input the target feature map into the detection head, and process the target feature map through the classification branch and regression branch of the detection head to obtain the category, bounding box and confidence of the target object in the original image.

[0052] A third aspect of the present invention provides an electronic device comprising a processor and a memory, wherein the memory stores at least one instruction, at least one program, a code set, or an instruction set, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by the processor to implement the small target detection method based on dynamic state space and flow matching as described in the first aspect.

[0053] A fourth aspect of the present invention provides a computer-readable storage medium storing at least one instruction, at least one program, a code set, or an instruction set, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to implement the small target detection method based on dynamic state space and flow matching as described in the first aspect.

[0054] The embodiments of the present invention have the following beneficial effects:

[0055] This invention provides a small target detection method based on dynamic state space and feature flow matching. The method includes: extracting multi-scale feature maps containing semantic and spatial information at different levels from the original image using a feature extraction module; inputting the multi-scale feature maps into a two-dimensional selective scanning module for depth processing to enhance key region features of the small target and suppress noise, resulting in an enhanced feature map that integrates dynamic spatial attention and multi-scale perturbation information; introducing a spatial dimension local transfer mechanism into the enhanced feature map and fusing and updating the state vector of the enhanced feature map to obtain a spatial-temporal enhanced feature map; mapping and converging the sparse and scattered feature distribution in the spatial-temporal enhanced feature map into a high-dimensional manifold region using a feature spatial flow matching mechanism based on Lie index mapping, resulting in a target feature map; inputting the target feature map into a detection head, processing the target feature map through the classification and regression branches of the detection head to obtain the category, bounding box, and confidence score of the target object in the original image. This scheme adds a multi-scale feature perturbation fusion branch, supplementing fine-grained feature information while preserving the original feature structure through multi-path feature perturbation and forced size alignment mechanisms. By dynamically updating the hidden state through a selective state-space mechanism, the ability to model long sequences is preserved while avoiding redundant computations of self-attention and the limitations of local receptive fields in traditional convolution, thus enabling on-demand allocation of computational resources. Stream matching data processing is introduced, and feature distribution is optimized through feature correction. Cross-channel attention is combined to enhance the response of key channels, and a scale-aware module is used to dynamically weight important regions such as small targets. With almost no increase in computational burden, the discriminative power of global features is improved, and the ability to capture local details is enhanced. Attached Figure Description

[0056] Figure 1 A flowchart illustrating the steps of a small target detection method based on dynamic state space and flow matching provided in this embodiment of the invention;

[0057] Figure 2 This is a structural block diagram of a small target detection device based on dynamic state space and flow matching, provided for an embodiment of the present invention. Detailed Implementation

[0058] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present solution, and not all embodiments. Based on the embodiments of the present solution, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present solution.

[0059] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of this disclosure, unless otherwise stated, "a plurality of" means two or more. Furthermore, the use of "based on" or "according to" implies openness and inclusiveness, because processes, steps, calculations, or other actions "based on" or "according to" one or more of the stated conditions or values ​​may in practice be based on additional conditions or beyond the stated values.

[0060] Figure 1 The flowchart illustrates the steps of a small target detection method based on dynamic state space and feature flow matching provided in this embodiment of the invention.

[0061] like Figure 1 As shown, the method includes:

[0062] Step 101: Extract multi-scale feature maps containing semantic and spatial information at different levels from the original image using the feature extraction module.

[0063] The original image to be used for target detection is acquired. This original image can be from a spacecraft's long-range exploration phase, and it contains "small targets" with extremely low pixel counts. This solution will perform target detection on these "small targets".

[0064] The original image is input into the feature extraction module (Backbone) of the network to extract multi-scale feature maps containing semantic information and spatial details at different levels.

[0065] Step 102: Input the multi-scale feature map into the two-dimensional selective scanning module for depth processing to enhance the key features of small targets and suppress noise, thereby obtaining an enhanced feature map that integrates dynamic spatial attention and multi-scale perturbation information.

[0066] The adaptive 2D selective scanning module (SS2D) performs depth processing on multi-scale feature maps to enhance the features of key regions such as small targets and suppress noise.

[0067] As an optional embodiment, step 102 includes:

[0068] Step 1021: Project the multi-scale feature map through the first convolutional layer and perform a nonlinear transformation through the first activation function to obtain the projected features.

[0069] The multi-scale feature maps are fed into the adaptive SS2D module. First, the input features are projected through a convolutional layer, and then a non-linear transformation is performed through an activation function to obtain the projected features.

[0070] Step 1022: The projected features are fed into the main feature processing branch and the multi-scale feature perturbation fusion branch in parallel for processing to obtain the main features and multi-scale perturbation features respectively;

[0071] Step 1023: Fuse the backbone features and the multi-scale perturbation features to obtain the fused features.

[0072] In steps 1022-1023, the projected features are fed in parallel into the main feature processing branch and the multi-scale feature perturbation fusion branch, the results of the two processes are fused, and the results are sent to the next step.

[0073] Step 1024: Input the fused features into the dynamic feature adjustment branch for processing to generate dynamic adjustment features.

[0074] The fused features are fed into the dynamic feature adjustment branch, and operations are performed to generate dynamically adjusted features.

[0075] Step 1025: Multiply the dynamically adjusted feature with the attention weight factor of the dynamic feature adjustment branch to obtain a weighted feature; send the weighted feature into the selective scanning module, perform sequence modeling through a parameterized state transition function, dynamically update the hidden state to capture long-distance dependencies between features, and obtain an enhanced feature map that integrates dynamic spatial attention and multi-scale perturbation information.

[0076] The adaptive SS2D module includes: a dynamic feature modulation branch, a multi-scale feature perturbation fusion branch, and a selective scan module. The selective scan module is the core component of the adaptive SS2D module.

[0077] Sequence modeling, in this paper, specifically refers to the way state-space models (SSMs) process image features. Sequence modeling flattens a 2D image into a 1D sequence, conveying information through "states" and capturing long-range dependencies.

[0078] Specifically, the dynamically adjusted features are multiplied by the weighting factor Z to obtain weighted features. These weighted features are then fed into the core selective scanning module, which performs sequence modeling through an efficient state-space mechanism, dynamically updating the hidden states to capture long-distance dependencies between features.

[0079] This sequence modeling process is implemented through a parameterized state transition function, specifically:

[0080]

[0081] in, For the first Time step, number Grouping, No. The state vector of the channel, σ( ) is the activation function. Here is the state transition matrix. loss function Regarding the first i The gradient at each time step.

[0082] The dynamic modulation operator is defined as follows:

[0083]

[0084] in, Indicates global average pooling. This represents the Hadamard product. This represents a two-dimensional convolution.

[0085] Finally, the selective scanning module outputs an "enhanced feature map" that incorporates dynamic spatial attention and multi-scale perturbation information.

[0086] As an optional embodiment, step 1022 includes:

[0087] Step 10221: Send the projected features into the main feature processing branch for processing to obtain the main features;

[0088] Step 10222: The projected features are fed in parallel into the first heterogeneous feature transformation path, the second heterogeneous feature transformation path, and the third heterogeneous feature transformation path of the multi-scale feature perturbation fusion branch;

[0089] Step 10223: Perform a linear transformation on the projected features through a 1×1 convolutional layer of the first heterogeneous feature transformation path to obtain the first feature;

[0090] Step 10224: Downsample the projected features by using a 3×3 depthwise convolution of the second heterogeneous feature transformation path, and after intermediate convolution processing, restore them to their original size by bilinear interpolation upsampling to obtain the second feature;

[0091] Step 10225: Compress the number of channels of the projected feature to C' through a 1×1 convolution of the third heterogeneous feature transformation path, then introduce spatial perturbation through a 3×3 depth convolution, and then restore the number of channels to C through a 1×1 convolution to obtain the third feature; where C is the original number of channels of the projected feature pair, and C' is the number of channels after compression.

[0092] Step 10226: Using a set of learnable weights, the first feature, the second feature, and the third feature are weighted and fused to obtain multi-scale perturbation features.

[0093] In steps 10221-10226, the backbone features are fed in parallel into three heterogeneous feature transformation paths with different structures.

[0094] The first heterogeneous feature transformation path performs a linear transformation on the features through a 1x1 convolution, keeping their spatial dimensions unchanged, in order to capture fine-grained details.

[0095] The second heterogeneous feature transformation path downsamples through a 3x3 depthwise convolution (with a stride of 2), and after intermediate convolution processing, it is upsampled through bilinear interpolation to restore the original size, thereby introducing scale perturbation.

[0096] The third heterogeneous feature transformation path compresses the number of channels to C' through a 1x1 convolution, introduces spatial perturbation through a 3x3 depthwise convolution, and then restores the number of channels to C through a 1x1 convolution.

[0097] Finally, the feature maps output from the three paths are weighted and fused using a set of learnable weights to obtain "multi-scale perturbation features," which enrich the diversity and hierarchy of the original features.

[0098] Here, "learnable weights" refers to parameters whose values ​​can be automatically adjusted and optimized through algorithms (such as backpropagation and gradient descent) during model training.

[0099] As an optional embodiment, step 1024 includes:

[0100] Step 10241: Input the fused feature into a 1×1 convolutional layer of the dynamic feature adjustment branch to compress the number of channels of the fused feature to obtain a compressed feature;

[0101] Step 10242: Input the compressed features into the activation function and perform a nonlinear transformation to obtain the activation features;

[0102] Step 10243: Input the activation feature into a 1×1 convolutional layer to restore the number of channels of the activation feature and obtain the restored feature;

[0103] Step 10244: Generate a spatial attention weight map with values ​​ranging from [0,1] by passing the recovered features through the Sigmoid function;

[0104] Step 10245: Multiply the spatial attention weight map and the projection feature element by element to obtain the dynamically adjusted feature.

[0105] In steps 10241-10245, the fused features are input into a 1x1 convolutional layer, and its number of channels is compressed to C / r (where C is the original number of channels and r is the compression ratio) to reduce the amount of computation.

[0106] The compressed features are input into the SiLU activation function and subjected to nonlinear transformation to obtain the activated features.

[0107] The activation features are input into another 1x1 convolutional layer, the number of channels is restored to C, and then a sigmoid function is passed through to generate a spatial attention weight map with values ​​in the range [0,1].

[0108] The learning process of this weight map can be achieved through the following modulation coefficient function. express:

[0109]

[0110] in, The weights and biases of the 1x1 compressed convolution in step 10241, To recover the weights and biases of the 1x1 convolution in step 10243, Here, u represents the sigmoid function, and u represents the input vector.

[0111] By multiplying the spatial attention weight map element-wise with the projected features, the key region features are dynamically enhanced and the irrelevant regions are suppressed, resulting in "dynamically adjusted features".

[0112] To quantify the effect of this adjustment, the channel-space response entropy is defined. ,when When the value is below the set threshold, it indicates that the module has achieved sparsity focusing of feature responses and increased the weight of key feature channels.

[0113] Step 103: Introduce a spatial dimension local transfer mechanism to the enhanced feature map and fuse and update the state vector of the enhanced feature map to obtain a spatial-temporal enhanced feature map.

[0114] This step aims to address the issue that the standard state-space model (SSM) in the adaptive SS2D module primarily focuses on temporal evolution but fails to adequately model the inherent spatial topology of the image. This step receives the "enhanced feature map" (containing the state vector from the previous time step) output from step 102 and updates the state vector by introducing a spatial dimension local transfer mechanism.

[0115] As an optional embodiment, step 103 includes:

[0116] Step 1031: The spatial dimension local transfer module receives the state vector of the enhanced feature map at the previous time step and the input features at the current time step.

[0117] Receive the state vector of the previous time step of the enhanced feature map and the input features at the current time .

[0118] Step 1032: Calculate the state of the previous state vector after time-series evolution based on the state transition function to obtain the time-series evolution state.

[0119] Regarding the state vector of the previous time step The following two computations can be performed in parallel:

[0120] First calculation method:

[0121] Temporal evolution calculation (standard SSM update): Calculate the state after temporal evolution according to the state transition function defined in step 1025.

[0122] Step 1033: Use the spatial transition operator to aggregate the spatial neighborhood information of the state vector at the previous time step, and use the spatial transition spectrum radius as a constraint index in the state update process to obtain the spatially aggregated state.

[0123] The second calculation method:

[0124] Spatial transfer computation (local neighborhood aggregation): applying the "spatial transfer operator" For the state vector of the previous time step Aggregate spatial neighborhood information.

[0125] This space transfer operator passes through a Learnable spatial convolution kernel K implementation:

[0126]

[0127] in, This is the spatial-temporal fusion coefficient, used to dynamically balance the contribution weights of spatial neighborhood information and temporal evolution information.

[0128] During model training, Adaptive optimization using gradient descent is employed when the input features contain rich spatial details (such as dense target scenes). It tends to increase the size to enhance spatial modeling capabilities; however, when features are primarily based on global temporal correlation (such as in large target scenes), This automatically reduces the time-series modeling advantage of SSM. From the perspective of manifold learning, the spatial transfer operator... This is equivalent to introducing geodesic corrections into the state space.

[0129] From the perspective of manifold learning, spatial transfer operators This is equivalent to introducing geodesic corrections into the state space. Let... and Let the exponential and logarithmic maps on the Riemannian manifold be respectively (used to describe the movement of points on the manifold), then the spatial transition process can be expressed as:

[0130]

[0131] The geometric meaning of this expression is: in the current state Starting from the point, move along the geodesic line on the manifold that points to the neighborhood aggregation result. The proportional distance allows for smooth transfer of spatial information while maintaining the topological structure of the state vector. This design avoids the feature distribution distortion that can occur with direct addition in Euclidean space, making it particularly suitable for spatial interaction modeling of high-dimensional visual features.

[0132] To ensure numerical stability of the newly introduced spatial computation process and avoid gradient explosion, this invention defines a spatial transfer spectrum radius. As a constraint indicator:

[0133]

[0134] Where s represents the spatial transition matrix, Let v represent the v-th eigenvalue of matrix S.

[0135] The spatial transfer spectral radius is a quantitative indicator that controls the stability of the spatial transfer kernel (3×3 convolution kernel). Its core function is to ensure the numerical stability of the local transfer process in the spatial dimension and avoid eigenvalue divergence by limiting the spectral radius (absolute value of the largest eigenvalue) of the convolution kernel to less than 1.

[0136] Theoretical analysis shows that when Time (of which) State transition matrix The spectral radius of the convolution kernel satisfies the Lyapunov stability condition during the state update process, thus avoiding gradient explosion or slow convergence. In practical implementation, this is achieved by adjusting the convolution kernel... Applying spectral normalization ensures It remains within a stable range. This mechanism is achieved through explicit modeling. Local spatial correlation enables the model to exhibit superior feature resolution in target edge detection tasks, especially in natural images with complex textures.

[0137] Step 1034: The temporal evolution state and the spatial aggregation state are weighted and fused using a learnable spatial-temporal fusion coefficient to generate the current state vector and obtain the spatial-temporal enhanced feature map.

[0138] Through a learnable "spatial-temporal fusion coefficient" The temporal evolution state is weighted and fused with the spatial transition state obtained in (b) to generate the final state vector at the current moment. .

[0139] Among them, the fusion coefficient Adaptive optimization is used during training to dynamically balance the contribution weights of spatial neighborhood information and temporal evolution information.

[0140] When dealing with visual tasks, traditional SSM models primarily focus on temporal state evolution in their core selective scanning process, while having limited ability to model the inherent spatial topology of images. Specifically, the state update of standard SSM relies solely on the linear transformation between the current input features and historical states, lacking explicit modeling of the correlations between spatially neighboring pixels. This results in insufficient feature representation capabilities in spatially sensitive areas such as target edges and texture details.

[0141] To address this deficiency, this invention proposes a spatial dimension local transfer mechanism, namely steps 1031-1034: First, a learnable spatial convolution kernel with a 3×3 neighborhood is defined, targeting the state dimension. Build spatial correlation modeling capabilities; then rearrange the spatiotemporal dimensions through dimensional rearrangement. By transforming it into a batch dimension, independent spatial convolution processing is achieved for features at each time step, effectively capturing... Local spatial relationships in the dimension; finally, the fusion ratio of spatial transfer features and the original state is controlled by a learnable weight factor α, while maintaining the linear complexity of SSM. At the same time, it overcomes the shortcomings of traditional scanning, which only focuses on temporal correlations and ignores spatial local correlations, and strengthens the spatial dependency modeling of small targets and their surrounding environment. By introducing spatial neighborhood interaction terms during the state update process, it realizes the dynamic evolution of the state vector in the spatial dimension.

[0142] The core idea of ​​this mechanism is to treat the state vector of each spatial location on the feature map as a point on a Riemannian manifold. By defining the geodesic distance on the manifold, the state of each location not only depends on its own historical evolution but also receives state transmissions from its spatial neighborhood (e.g., within a 3×3 range), thereby capturing local spatial correlations. Specifically, the modified state update equation is as follows:

[0143]

[0144] in, and Let A represent the state vectors at time steps t and t-1, respectively, and let A represent the system transition matrix. B represents the time difference between two time steps, and B represents the input matrix. This represents the external input signal at time t.

[0145] As a spatial transition operator, its physical meaning is the state vector of the previous time step. Spatial neighborhood information aggregation is performed. This operator uses a 3×3 learnable convolution kernel. Implementation, specifically defined .

[0146] Step 104: Through the feature space flow matching mechanism based on Lie index mapping, the sparse and scattered feature distribution in the spatial-temporal enhanced feature map is mapped and converged into a high-dimensional manifold region to obtain the target feature map.

[0147] This step aims to address the core issues of scattered, blurred, and confused features of small targets with background features in single-frame images. This step receives the "spatial-temporal enhanced feature map" output from step 103, and through a feature space flow matching mechanism based on Lie Group Exponential Mapping, maps and converges the originally sparse and scattered feature distribution of small targets into a more compact and discriminative high-dimensional manifold region to obtain the target feature map.

[0148] As an optional embodiment, step 104 includes:

[0149] Step 1041: Obtain the original distribution of the single-frame small target feature set in the spatial-temporal enhanced feature map, select a special Euclidean group as a Lie group, and map the original distribution to the compact discriminative distribution of large target features through Lie group transformation to obtain the transformed distribution.

[0150] First, a Lie group representation of the feature distribution is performed. Specifically, the set of small target features in a single frame from the feature map output in step 103 is set as... (in for (dimensional feature vector), whose original distribution is Selecting a special Euclidean group As Li Qun Its corresponding Lie algebra Used to characterize continuous transformations in the feature space.

[0151] Then, distribution alignment based on the Lie exponent mapping is performed. The goal is to achieve distribution alignment through Lie group transformation. Compact discriminative distribution mapped to large target features This transformation is achieved through the Lie exponent mapping. accomplish:

[0152]

[0153] in, It is a Lie algebra element representing the "velocity field" of the transformation. This is the Lie exponent mapping, which transforms discrete Lie algebra elements into continuous transformations over the Lie group.

[0154] "Velocity field" The goal, learned through learning, is to minimize the mean vector of the transformed features and the large target features. Distance between:

[0155]

[0156] In the formula The mean vector of the large target features is used to ensure that the small target features are mapped to the discriminative region by minimizing the distance between the transformed features and the center of the large target. This represents the original distribution.

[0157] Step 1042: Utilize Lie group action to achieve collaborative optimization of the cross-scale features of the transformed distribution, complete the fracture features of small targets, and obtain optimized distribution features.

[0158] This process utilizes Lie group action to achieve collaborative optimization of cross-scale features in order to complete the fracture features of small targets.

[0159] As an optional embodiment, step 1042 includes:

[0160] Step 10421: Define the transformation from high-scale features to low-scale features as a Lie algebra mapping.

[0161] The model needs to know how to compare "high-scale features" (rich in detail) with "low-scale features" (semantically rich but less detailed). To do this, it defines a Lie algebra mapping. This mapping acts like a "translator" or "velocity field," describing how to transform and align high-scale features to low-scale features.

[0162] Define from high-scale features to low-scale features The transformation is a Lie algebra mapping. :

[0163]

[0164] in This is the scaling transformation matrix (achieved through convolution and interpolation). The Lie logarithmic map (the inverse operation of the Lie exponential map) transforms a discrete scaling transformation into a continuous Lie algebraic "velocity field". s and s ‘ This indicates two different scales.

[0165] Step 10422: Using the Lie algebra mapping, calculate the distances of features at different scales on the Lie group manifold to obtain the Lie group distance.

[0166] During training, the model uses the mapping defined in the first step to calculate the "distance" between features at different scales on the Lie group manifold. The method for calculating the Lie group distance is as follows:

[0167]

[0168] in, This indicates a transformation from scale s to scale s. ‘ Lie group elements, This represents the features extracted at scale s. Representation of features The inverse element in a Lie group.

[0169] Step 10423: Calculate the Lie group entropy of the feature distribution at each scale, and determine the fusion weight of each Lie group entropy based on the principle that the smaller the Lie group entropy, the higher the weight; according to the fusion weight, weightedly fuse the features at each scale, and in the weighted fusion process, minimize the Lie group distance to constrain the consistency of features at different scales in the fusion process to obtain the optimized distribution features.

[0170] The fusion weights of multi-scale features are determined by the "Lie group entropy" of the features. Decision, integration weight The calculation method is as follows:

[0171]

[0172] in, Let the entropy of the Lie group for the s-th feature be denoted as . Let represent the probability density function of the s-th feature on the Lie group G.

[0173] In the weighted fusion process, the consistency of features at different scales is constrained by minimizing the Lie group distance. This constraint reduces the "path differences" of features at different scales on the Lie group, enhancing the cross-scale continuity of small target features. It utilizes cross-scale information (i.e., introducing features from other scales) to "fill in" missing or broken information at the current scale. Through the Lie group distance constraint, it ensures that features at different scales are consistent, thereby "recovering" lost details.

[0174] Step 1043: Apply a Lie algebra regularization term to the optimized distribution feature, the Lie algebra regularization term being obtained based on the gradient on the Lie group manifold; by minimizing the regularization term, reduce the jumps of the optimized distribution feature on the high-dimensional manifold, and obtain the target feature map.

[0175] To address the issue of edge and texture "fragmentation" caused by insufficient sampling of small targets, feature smoothing is achieved through gradient descent on Lie group manifolds.

[0176] For optimizing distribution characteristics Apply a Lie algebra regularization term :

[0177]

[0178] in, Features The gradient on a Lie group manifold is defined as:

[0179] Where is an infinitesimal quantity approaching 0. Denotes the exponent of the Lie algebra to the Lie group.

[0180] Mapping.

[0181] By minimizing this regularization term, the abrupt changes of features on high-dimensional manifolds are reduced, making the edge and texture features of small objects smoother and more continuous.

[0182] This step outputs the "target feature map" after the above distribution alignment, multi-scale collaborative optimization, and manifold smoothing processes. This target feature map has increased the distance between small targets and background features in the feature space, improving the separability of features.

[0183] Step 105: Input the target feature map into the detection head, and process the target feature map through the classification branch and regression branch of the detection head to obtain the category, bounding box and confidence of the target object in the original image.

[0184] This step receives the target feature map output from step 104. This target feature map integrates all the enhancement information from temporal, spatial, multi-scale, and manifold matching, and has a high degree of separability.

[0185] The target feature map is input into the detection head. The detection head processes the target feature map through its internal parallel network branches, which include a classification branch and a regression branch.

[0186] Specifically, the classification branch calculates the category (e.g., "spacecraft component" or "background") of each potential target in the feature map through the classification network and generates the corresponding confidence score.

[0187] The regression branch calculates the bounding box coordinates of each potential target using a regression network.

[0188] The classification branch outputs the category information and confidence score, and the regression branch outputs the bounding box coordinates. Finally, the combined information (bounding box, category, confidence score) is output as the final small object detection result.

[0189] In summary, the main tasks of this plan are as follows:

[0190] 1. Two-dimensional selective scanning module for channel space integration: To solve the size mismatch problem that easily occurs in the multi-scale feature processing of existing scanning modules, and to enhance the diversity and robustness of small target features, this solution adds a multi-scale feature perturbation fusion branch. Through multi-path feature perturbation and forced size alignment mechanism, fine-grained feature information is supplemented while preserving the original feature structure.

[0191] 2. Efficient global-local collaborative modeling: Inheriting the linear complexity advantage of the Mamba architecture, it dynamically updates the hidden state through a selective state space mechanism. While retaining the ability to model long sequences, it avoids the redundant computation of Transformer self-attention and the local receptive field limitation of traditional convolution. In the feature extraction process, it achieves on-demand allocation of computing resources by adjusting the scanning range with scene awareness (dynamic adaptation of stride and window size) to "redundancy reduction in simple scenes and accuracy improvement in complex scenes".

[0192] 3. By introducing streaming matching data processing, the model optimizes the feature distribution through feature correction, enhances the response of key channels by combining cross-channel attention, and dynamically weights important regions such as small targets with the help of the scale awareness module. Under the premise of almost no increase in computational burden, it can improve the discriminativeness of global features and strengthen the ability to capture local details. At the same time, it is compatible with the original architecture through residual connection, effectively enhancing the robustness and accuracy of the model in complex scenarios.

[0193] Compared to existing methods, the advantages of this invention's algorithm are: stronger global correlation modeling capabilities than the YOLO series, and significant optimization in hardware resource consumption. By simplifying redundant computation modules and optimizing feature processing, the number of parameters and computational cost (FLOPs) can be greatly reduced, saving a significant amount of hardware resources while ensuring improved accuracy, making it more suitable for resource-constrained scenarios such as edge computing; it is also more efficient than Transformer-based methods; it has more flexible scene adaptability than Mamba YOLO, with a dynamic scanning strategy that can specifically enhance the capture of details of small targets, and this dynamic strategy does not require high-resolution processing of the entire scene, further reducing the unnecessary occupation of hardware resources; the combination of flow matching and SSM provides an end-to-end solution for small target detection from feature optimization to distribution alignment, avoiding the additional resource consumption caused by multi-stage post-processing, and further enhancing the advantage of efficient utilization of hardware resources.

[0194] This invention focuses on the bottleneck of small target detection during long-range spacecraft exploration. To address the contradiction between the locality of traditional CNNs and the high complexity of Transformers, it proactively explores the application potential of state-space models. This method achieves dynamic focusing of computational resources towards high-information regions by designing an adaptive SS2D module; it introduces a spatial dimension local transfer mechanism to enhance the ability to capture spatial details of small targets; and it constructs a feature space flow matching mechanism to optimize the separability of single-frame small target features without requiring temporal information. Experiments demonstrate that this method improves the AP by 6.8% compared to the baseline model on a subset of long-range small targets in the SatKeyComps-DS dataset, validating its advantage in efficiently modeling weak target features with linear complexity.

[0195] In summary, the small target detection method based on dynamic state space and feature flow matching provided by this invention includes: extracting multi-scale feature maps containing semantic and spatial information at different levels from the original image through a feature extraction module; inputting the multi-scale feature maps into a two-dimensional selective scanning module for depth processing to enhance key region features of small targets and suppress noise, resulting in an enhanced feature map that integrates dynamic spatial attention and multi-scale perturbation information; introducing a spatial dimension local transfer mechanism into the enhanced feature map and fusing and updating the state vector of the enhanced feature map to obtain a spatial-temporal enhanced feature map; mapping and converging the sparse and scattered feature distribution in the spatial-temporal enhanced feature map into a high-dimensional manifold region through a feature spatial flow matching mechanism based on Lie index mapping to obtain a target feature map; inputting the target feature map into a detection head, and processing the target feature map through the classification and regression branches of the detection head to obtain the category, bounding box, and confidence of the target object in the original image. This scheme adds a multi-scale feature perturbation fusion branch, which supplements fine-grained feature information while preserving the original feature structure through multi-path feature perturbation and forced size alignment mechanisms. By dynamically updating the hidden state through a selective state-space mechanism, the ability to model long sequences is preserved while avoiding redundant computations of self-attention and the limitations of local receptive fields in traditional convolution, thus enabling on-demand allocation of computational resources. Stream matching data processing is introduced, and feature distribution is optimized through feature correction. Cross-channel attention is combined to enhance the response of key channels, and a scale-aware module is used to dynamically weight important regions such as small targets. With almost no increase in computational burden, the discriminative power of global features is improved, and the ability to capture local details is enhanced.

[0196] Figure 2 This is a structural block diagram of a small target detection device based on dynamic state space and flow matching provided in an embodiment of the present invention. Figure 2 As shown, the device 200 includes:

[0197] The feature extraction module 201 is used to extract multi-scale feature maps containing semantic and spatial information at different levels from the original image.

[0198] The depth processing module 202 is used to input the multi-scale feature map into the two-dimensional selective scanning module for depth processing, enhance the key features of small targets and suppress noise, and obtain an enhanced feature map that integrates dynamic spatial attention and multi-scale perturbation information.

[0199] The fusion update module 203 is used to introduce a spatial dimension local transfer mechanism into the enhanced feature map and to fuse and update the state vector of the enhanced feature map to obtain a spatial-temporal enhanced feature map.

[0200] The aggregation module 204 is used to map and aggregate the sparse and scattered feature distributions in the spatial-temporal enhanced feature map into a high-dimensional manifold region through a feature space flow matching mechanism based on Lie index mapping, thereby obtaining the target feature map;

[0201] The detection module 205 is used to input the target feature map into the detection head, and process the target feature map through the classification branch and regression branch of the detection head to obtain the category, bounding box and confidence of the target object in the original image.

[0202] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0203] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0204] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that comply with the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the following claims.

[0205] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. A small target detection method based on dynamic state space and feature flow matching, characterized in that, The method includes: The original image is processed by a feature extraction module to extract multi-scale feature maps containing semantic and spatial information at different levels; The multi-scale feature map is input into the two-dimensional selective scanning module for depth processing to enhance the key features of small targets and suppress noise, resulting in an enhanced feature map that integrates dynamic spatial attention and multi-scale perturbation information. A spatial dimension local transfer mechanism is introduced into the enhanced feature map, and the state vector of the enhanced feature map is fused and updated to obtain a spatial-temporal enhanced feature map; By using a feature space flow matching mechanism based on Lie index mapping, the sparse and scattered feature distributions in the spatial-temporal enhanced feature map are mapped and converged into a high-dimensional manifold region to obtain the target feature map; The target feature map is input into the detection head, and the target feature map is processed by the classification branch and regression branch of the detection head to obtain the category, bounding box and confidence of the target object in the original image.

2. The method according to claim 1, characterized in that, The step of inputting the multi-scale feature map into a two-dimensional selective scanning module for depth processing to enhance the features of key regions of small targets and suppress noise includes: The multi-scale feature map is projected through the first convolutional layer and then nonlinearly transformed using the first activation function to obtain the projected features. The projected features are fed in parallel into the main feature processing branch and the multi-scale feature perturbation fusion branch for processing, to obtain the main features and multi-scale perturbation features respectively. The core features and the multi-scale perturbation features are fused to obtain the fused features; The fused features are input into the dynamic feature adjustment branch for processing to generate dynamically adjusted features; The dynamically adjusted features are multiplied by the attention weight factor of the dynamically adjusted feature branch to obtain weighted features. The weighted features are then fed into the selective scanning module, where sequence modeling is performed using a parameterized state transition function. The hidden states are dynamically updated to capture long-distance dependencies between features, resulting in an enhanced feature map that integrates dynamic spatial attention and multi-scale perturbation information.

3. The method according to claim 2, characterized in that, The process of feeding the projected features in parallel into the main feature processing branch and the multi-scale feature perturbation fusion branch to obtain the main features and multi-scale perturbation features respectively includes: The projected features are sent to the main feature processing branch for processing to obtain the main features. The projected features are fed in parallel into the first heterogeneous feature transformation path, the second heterogeneous feature transformation path, and the third heterogeneous feature transformation path of the multi-scale feature perturbation fusion branch; The first feature is obtained by linearly transforming the projected feature through a 1×1 convolutional layer of the first heterogeneous feature transformation path; The projected features are downsampled by a 3×3 depthwise convolution of the second heterogeneous feature transformation path, and after intermediate convolution processing, they are upsampled by bilinear interpolation to restore them to their original size, thus obtaining the second feature. The number of channels of the projected feature is compressed to C' through a 1×1 convolution of the third heterogeneous feature transformation path, then spatial perturbation is introduced through a 3×3 depthwise convolution, and then the number of channels is restored to C through a 1×1 convolution to obtain the third feature; where C is the original number of channels of the projected feature pair, and C' is the number of channels after compression. The first feature, the second feature, and the third feature are weighted and fused using a set of learnable weights to obtain multi-scale perturbation features.

4. The method according to claim 2, characterized in that, The fused features are input into the dynamic feature adjustment branch for processing to generate dynamically adjusted features, including: The fused features are input into a 1×1 convolutional layer of the dynamic feature adjustment branch to compress the number of channels of the fused features, thus obtaining compressed features. The compressed features are input into the activation function and subjected to nonlinear transformation to obtain the activation features; The activation features are input into a 1×1 convolutional layer to restore the number of channels of the activation features, thus obtaining the restored features; The recovered features are processed by the Sigmoid function to generate a spatial attention weight map with values ​​ranging from [0,1]. The spatial attention weight map is multiplied element-wise with the projected features to obtain dynamically adjusted features.

5. The method according to claim 1, characterized in that, The process of introducing a spatial dimension local transfer mechanism into the enhanced feature map and fusing and updating the state vector of the enhanced feature map to obtain a spatial-temporal enhanced feature map includes: The spatial dimension local transfer module receives the previous state vector and the current input features of the enhanced feature map; The state after the temporal evolution of the state vector at the previous moment is calculated based on the state transition function to obtain the temporal evolution state; The spatial neighborhood information of the state vector at the previous time step is aggregated using the spatial transition operator, and the spatial transition spectrum radius is used as a constraint index in the state update process to obtain the spatially aggregated state. The temporal evolution state and the spatial aggregation state are weighted and fused using a learnable spatial-temporal fusion coefficient to generate the current state vector and obtain the spatial-temporal enhanced feature map.

6. The method according to claim 1, characterized in that, The feature space flow matching mechanism based on Lie index mapping maps and converges the sparse and scattered feature distributions in the spatial-temporal enhanced feature map into a high-dimensional manifold region to obtain the target feature map, including: Obtain the original distribution of the single-frame small target feature set in the spatial-temporal enhanced feature map, and select a special Euclidean group as a Lie group. Map the original distribution to the compact discriminative distribution of large target features through Lie group transformation to obtain the transformed distribution. By utilizing Lie group operations, the cross-scale features of the transformed distribution are collaboratively optimized to complete the fracture features of small targets and obtain optimized distribution features. A Lie algebra regularization term is applied to the optimized distribution feature, which is obtained based on the gradient on the Lie group manifold; by minimizing the regularization term, the jumps of the optimized distribution feature on the high-dimensional manifold are reduced, and the target feature map is obtained.

7. The method according to claim 6, characterized in that, The method of utilizing Lie group interactions to achieve collaborative optimization of the cross-scale features of the transformed distribution, completing the fracture features of small targets, and obtaining optimized distribution features includes: The transformation from high-scale features to low-scale features is defined as a Lie algebra mapping; Using the Lie algebra mapping, the distances of features at different scales on the Lie group manifold are calculated to obtain the Lie group distance; Calculate the Lie group entropy of the feature distribution at each scale, and determine the fusion weight of each Lie group entropy based on the principle that the smaller the Lie group entropy, the higher the weight; according to the fusion weight, the features at each scale are weighted and fused, and in the process of weighted fusion, the consistency of features at different scales is constrained by minimizing the Lie group distance during the fusion process to obtain the optimized distribution features.

8. A small target detection device based on dynamic state space and feature flow matching, characterized in that, The device includes: The feature extraction module is used to extract multi-scale feature maps containing semantic and spatial information at different levels from the original image. The depth processing module is used to input the multi-scale feature map into the two-dimensional selective scanning module for depth processing, enhance the key features of small targets and suppress noise, and obtain an enhanced feature map that integrates dynamic spatial attention and multi-scale perturbation information. The fusion update module is used to introduce a spatial dimension local transfer mechanism into the enhanced feature map and to fuse and update the state vector of the enhanced feature map to obtain a spatial-temporal enhanced feature map. The aggregation module is used to map and aggregate the sparse and scattered feature distributions in the spatial-temporal enhanced feature map into a high-dimensional manifold region through a feature space flow matching mechanism based on Lie index mapping, thereby obtaining the target feature map; The detection module is used to input the target feature map into the detection head, and process the target feature map through the classification branch and regression branch of the detection head to obtain the category, bounding box and confidence of the target object in the original image.

9. An electronic device, characterized in that, The electronic device includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, a code set, or an instruction set, and the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by the processor to implement the small target detection method based on dynamic state space and flow matching as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, the at least one program, the code set, or instruction set is loaded and executed by a processor to implement the small target detection method based on dynamic state space and flow matching as described in any one of claims 1-7.

Images