A method for fast target detection in space

Abstract

The application relates to the technical field of target detection, and specifically provides a space fast target detection method, which comprises the following steps: an improved U-Net architecture is obtained by improving a traditional U-Net architecture, the improved U-Net architecture is used as a backbone network, an improved feature extraction network is constructed, the improved method of the improved U-Net architecture is that, in an encoder part of the traditional U-Net architecture, a standard convolution layer in a basic construction unit is replaced by a direction self-adaptive rectangular convolution module, a total loss function of the improved feature extraction network is a loss function mixed by physical perception and skeleton anchoring, the improved feature extraction network is trained, and the trained improved feature extraction network is used to detect a space fast target. The application constructs a "rotating rectangular" sampling grid highly consistent with the physical form of the space fast target, accurately captures the space fast targets with different characteristics, and improves the detection precision.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and specifically provides a method for rapid spatial target detection. Background Technology

[0002] Fast target detection in space is a key technology in astronomical image processing and space situational awareness. Fast targets in space typically exhibit high relative speeds. During imaging detection, due to the limited exposure time of the detector, the target spans multiple pixel positions within a single frame, causing its energy to expand and accumulate along the direction of motion. Therefore, these targets often no longer appear as ideal point-like or compact speckle shapes in space target images, but rather as stripes, trails, or lines with distinct directions. This imaging characteristic reflects the motion blur effect produced by the combined effect of high-speed target motion and limited exposure time. Furthermore, the background of images containing fast targets in space is complex, containing various interference sources such as stellar point sources, cosmic ray impacts, thermal noise, readout noise, and optical stray light. The signal-to-noise ratio between fast targets and background noise is often low, posing a severe challenge to the robustness of detection algorithms.

[0003] In terms of traditional methods, early research mainly relied on geometric transformations and template matching techniques. For example, line detection methods based on Radon and Hough transforms identify linear structures by mapping the image space to the parameter space and performing cumulative voting; directional template matching methods use multiple pre-set filters at different angles to perform convolutional response analysis on the image; in addition, there are multi-stage hypothesis testing and pattern clustering. These methods can achieve stripe detection under ideal conditions, but they generally suffer from problems such as heavy reliance on pre-set parameters and thresholds, sensitivity to scene changes, and poor generalization ability. When faced with complex situations such as noise interference and stripe breakage and adhesion, they are prone to false alarms and missed detections, and the computational complexity increases sharply with the dimension of the parameter space, making it difficult to meet the real-time processing requirements of large-scale data.

[0004] In recent years, deep learning methods, represented by convolutional neural networks, have been introduced into this task. Leveraging their advantages of end-to-end learning and automatic feature extraction, their overall performance significantly outperforms traditional methods. However, existing deep learning methods still exhibit several shortcomings when dealing with fast spatial targets. First, the standard convolution of mainstream networks uses a square sampling window, implicitly assuming isotropy. This fails to match the anisotropic geometric features of stripes extending long distances along the principal axis and narrow, limited variation along the normal, resulting in insufficient representation of slender structures and easily broken stripes. Second, while existing adaptive convolution methods such as ARConv can dynamically adjust the sampling grid size, they only support scaling along the horizontal and vertical axes, failing to achieve arbitrary angle rotation and alignment, and severely lacking geometric adaptability to diagonal stripes. Furthermore, although deformable convolution breaks through the limitation of fixed sampling positions, the lack of physical prior constraints on the stripe structure makes sampling offsets prone to following background noise, leading to overfitting in low signal-to-noise ratio scenarios, resulting in increased false alarms and distorted target shapes. Finally, existing loss functions mostly optimize for pixel-level classification errors, failing to suppress boundary label uncertainty caused by motion blur and unable to explicitly constrain key physical morphology aspects such as centroid localization accuracy, principal direction consistency, and linearity. Prediction results often exhibit topological and morphological errors such as breaks, bends, adhesions, or degradation into spots. Furthermore, existing small target detection loss functions, such as SLS Loss and SD Loss, are primarily designed for point-like or patchy targets. Their scale weighting based on area ratios and constraint only on centroid position-sensitive terms make them ill-suited for stripe structures with extremely large aspect ratios. They cannot effectively distinguish between short, thick targets and long, thin targets, nor can they guarantee accurate stripe boundary localization. In summary, existing technologies have significant shortcomings in geometric adaptability, noise robustness, and morphological consistency constraints, necessitating a novel detection method that balances adaptive feature extraction with both orientation and scale, as well as physical geometric constraints. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a method for rapid spatial target detection. It introduces an orientation-adaptive rectangular convolution module into an improved feature extraction network and constructs a total loss function that combines physical perception and skeleton anchoring. This enables accurate detection and localization of rapid spatial targets, improving generalization ability and efficiency.

[0006] The present invention provides a fast spatial target detection method, comprising: S1: Improve the traditional U-Net architecture to obtain the improved U-Net architecture, and use the improved U-Net architecture as the backbone network to build an improved feature extraction network; The improvement method of the U-Net improved architecture is as follows: in the encoder part of the traditional U-Net architecture, the standard convolutional layer in the basic building unit is replaced with an orientation-adaptive rectangular convolutional module. The total loss function of the improved feature extraction network is: ; in, This represents the skeleton anchoring weighted loss. This represents the loss of geometric consistency during decoupling. Indicates the dynamic weighting coefficient; S2: Training an improved feature extraction network; S3: Utilize the improved feature extraction network after training to detect spatial targets quickly.

[0007] Preferably, the orientation-adaptive rectangular convolution module includes a length prediction subnetwork and an orientation prediction subnetwork, which are used to calculate the length field and orientation field of the prediction convolution kernel from the input feature map, respectively.

[0008] Preferably, the orientation-adaptive rectangular convolution module includes the following sub-steps: Sub-step 1: Utilize the length prediction sub-network and the orientation prediction sub-network to solve the input feature map pixel by pixel, and predict the length field and orientation field of the effective convolutional kernel at each pixel; Sub-step 2: Dynamically generate a rectangular sampling grid at each pixel position based on the length field and orientation field. The rectangular sampling grid is stretched along the tangent vector direction and maintains a specific width along the normal vector direction. Calculate the actual position using the sampling points corresponding to the size of the effective convolution kernel. Sub-step 3: Perform bilinear interpolation on the actual location to generate a resampled feature map; Sub-step 4: Preset a set of widths, and perform sub-step 2 and sub-step 3 on each width in the width set to obtain multiple resampled feature maps; Perform a convolution operation on the resampled feature map to obtain branch features; The final feature is obtained by weighted summation of the branch features.

[0009] Preferably, the sub-steps of the orientation-adaptive rectangular convolution module further include: Sub-step 5: Apply an affine transformation to the final features to obtain the encoder's output features. , is represented as: ; in, This indicates that the final feature is obtained by weighted summation of the branch features. This represents element-wise multiplication. This represents element-wise addition. and Both represent matrices for predicting affine transformations.

[0010] Preferably, each pixel Length of the effective convolution kernel at the point Represented as: ; in, Indicates the length scaling factor. This represents the bias factor. This represents a single-channel scalar plot. To predict from the input feature map using a length prediction subnetwork; The orientation prediction subnetwork predicts a two-channel vector map from the input feature map, and the two-channel vector map is L2 normalized to obtain a unit tangent vector field representing the principal axis directions of the fast spatial target. and the unit normal vector field orthogonal to it , is represented as: ; ; in, Represents a two-channel vector graph. This represents a small constant used to prevent division by zero. This represents the x-direction component of the unit tangent vector. This represents the y-direction component of the unit tangent vector.

[0011] The preferred method for calculating the actual location is as follows: ; in, This represents the center pixel coordinates of the effective convolution kernel. This represents the index of the sampling point corresponding to the effective convolution kernel size. Represents normalized local relative coordinates, , Indicates the effective kernel length. This indicates the kernel width set for the current branch.

[0012] Preferred, resampled feature map Represented as: ; in, Indicates the actual location. Indicates the input feature map, This represents the bilinear interpolation operator.

[0013] Preferred, final features Represented as: ; in, Indicates branching characteristics, This represents the pixel-level weight map predicted by the gated subnetwork. Indicates a branch of width.

[0014] Preferably, after obtaining the output features, the output features enter the decoder, which is connected to the MSHNet multi-scale prediction head. The multi-scale prediction head generates prediction maps at different resolution levels of the decoder, and then upsamples the prediction maps to the size of the input feature map and fuses them to obtain a stripe probability map.

[0015] Preferably, the calculation process of the total loss function is divided into three parts: Part 1: Calculating the Weighted Loss of Skeleton Anchoring : ; in, Represents the normalization factor. This represents the two-dimensional coordinates of the corresponding pixel in the prediction probability map, the ground truth mask, and the spatial weight map. Indicates the predicted probability. Indicates the focus parameter, Represents a spatial weighted graph; Part Two: Calculating the Decoupling Geometric Consistency Loss : , in, This represents the centroid of the prediction probability graph. Denotes the centroid of the truth graph. This represents a small constant used to prevent division by zero. Indicates the weighting coefficient. This indicates the weight of the corresponding item. This represents the first principal eigenvector of the covariance matrix corresponding to the predicted probability map. This represents the first principal eigenvector of the covariance matrix corresponding to the truth map. This represents the first eigenvalue of the covariance matrix corresponding to the predicted probability map. This represents the second eigenvalue of the covariance matrix corresponding to the predicted probability map. This represents the first eigenvalue of the covariance matrix corresponding to the truth map. This represents the second eigenvalue of the covariance matrix corresponding to the truth map. Indicates the centroid location term. Indicates the consistency term in the main direction. Represents the linearity term; Part Three: Adjusting Dynamic Course Learning Using Dynamic Weighting Coefficients. The dynamic weighting coefficients are expressed as follows: ; in, This represents the average intersection-union ratio (IoU) between the predicted probability map and the ground truth mask in the current training batch. The hyperparameters that control the steepness of the Sigmoid curve are... This indicates the preset crossover ratio threshold.

[0016] Compared with the prior art, the present invention can achieve the following beneficial effects: Compared with existing technologies, the spatial fast target detection method proposed in this invention introduces orientation-adaptive rectangular convolution into the improved feature extraction network. By explicitly predicting the fast length field and orientation field, it constructs a "rotating rectangle" sampling grid that highly matches the physical shape of the spatial fast target, accurately captures anisotropic features, enhances feature continuity, and at the same time avoids sampling points from diverging into clutter through geometric constraints, thereby improving recall and significantly reducing false alarm rate.

[0017] The improved feature extraction network of this invention features a skeleton-anchored weighting mechanism in its total loss function. It utilizes Euclidean distance transformation to construct a Gaussian weight field centered on the skeleton, enabling the improved feature extraction network to prioritize high-confidence centerline regions, reducing sensitivity to blurred edges and significantly improving tolerance to label noise and localization accuracy. The total loss function incorporates decoupled geometric consistency constraints. Using second-order moment eigenvalue decomposition, it independently incorporates three physical attributes—centroid position, principal direction consistency, and linearity (aspect ratio)—into the optimization objective, forcing the network output to conform to physical laws and ensuring the integrity of the detection results in terms of topology and geometry.

[0018] In addition, the present invention integrates a direction-adaptive rectangular convolution with multi-width parallel branches and a gating fusion mechanism, dynamically allocates weights, and automatically adjusts the receptive field width. It can simultaneously detect fine debris trajectories and coarser satellite trajectories in the same model without adjusting the network structure, thus significantly improving generalization ability and efficiency. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the execution process of the orientation-adaptive rectangular convolution module provided in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and do not constitute a limitation thereof. Similar elements in different embodiments are referred to by associated similar element reference numerals. In the following embodiments, many details are described to facilitate a better understanding of the invention. However, those skilled in the art will readily recognize that some features may be omitted in different situations, or may be replaced by other elements, materials, or methods. In some cases, some operations related to the invention are not shown or described in the specification. This is to avoid obscuring the core parts of the invention with excessive description. For those skilled in the art, detailed description of these related operations is not necessary; they can fully understand the related operations based on the description in the specification and general technical knowledge in the art.

[0021] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined to form various implementations. Furthermore, the order of the steps or actions in the method description can be changed or adjusted in a manner readily apparent to those skilled in the art. Therefore, the various orders in the specification and drawings are merely for the clear description of a particular embodiment and do not imply a mandatory order, unless otherwise stated that a particular order must be followed.

[0022] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation on this invention. Furthermore, the terms "first," "second," etc., 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, features defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0023] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0024] The invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0025] This invention provides a method for fast target detection in space, comprising: S1: Improve the traditional U-Net architecture to obtain an improved U-Net architecture. Use the improved U-Net architecture as the backbone network to construct an improved feature extraction network. The improvement method of the U-Net architecture is as follows: in the encoder part of the traditional U-Net architecture, replace the standard convolutional layers in the basic building units with orientation-adaptive rectangular convolutional modules. The total loss function of the improved feature extraction network is... .

[0026] In this embodiment of the invention, an improved feature extraction network is constructed for fast spatial target detection. The backbone network of the improved feature extraction network is an improved U-Net architecture, which is obtained by improving upon the traditional U-Net architecture. The traditional U-Net architecture includes an encoder, a decoder, and a multi-scale prediction head. The encoder is used to extract abstract semantic features of the image step by step. The decoder is used to recover the spatial resolution of the feature map. The decoder fuses the features of the encoder through skip connections and uses the multi-scale head to output prediction results at different resolution levels.

[0027] This invention retains the U-shaped structure and multi-scale prediction head of the traditional U-Net architecture, while improving its encoder structure. In each downsampling stage of the traditional U-Net encoder, there are typically consecutive 3×3 standard convolutional layers or residual blocks. This invention replaces the standard convolutional layers in these basic building blocks with Oriented Adaptive Rectangular Convolution (OARConv) modules to obtain an improved U-Net architecture. This improved U-Net architecture is then used as the backbone network of an improved feature extraction network, which is essentially a deep neural network.

[0028] like Figure 1As shown, the orientation-adaptive rectangular convolution module aims to dynamically adjust the shape and orientation of the effective convolution kernel based on the local geometric features of the spatial target. The orientation-adaptive rectangular convolution module includes a length prediction subnetwork and an orientation prediction subnetwork, which are used to compute the length field and orientation field of the predicted convolution kernel from the input feature map, respectively. This invention embodiment uses... Indicates the input feature map, ,in, and These represent the height and width of the input feature map, respectively. This indicates the number of input channels. The execution process of the orientation-adaptive rectangular convolution module includes the following sub-steps: Sub-step 1: Utilize the length prediction sub-network and the orientation prediction sub-network to solve the input feature map pixel by pixel, predicting the length field and orientation field of the effective convolutional kernel at each pixel.

[0029] The orientation-adaptive rectangular convolution module internally contains two parallel lightweight convolutional sub-networks: a length prediction sub-network and an orientation prediction sub-network. These are used to compute the length and orientation fields of the effective convolutional kernels pixel-by-pixel from the input feature map, respectively. Specifically, the length prediction sub-network utilizes... From the input feature map The prediction yields a single-channel scalar map. In order to obtain the single-channel scalar map from the prediction With the value constrained within a reasonable range, this embodiment of the invention uses the Sigmoid activation function and a linear scaling formula to calculate the value of each pixel. Length of the effective convolution kernel at the point , is represented as: , in, This represents the length scaling factor, used to set the maximum receptive field range. This represents the bias factor, which guarantees the minimum convolution kernel length.

[0030] Using direction prediction subnetwork From the input feature map Predicting dual-channel vector graph , L2 normalization is performed on the dual-channel vector map to obtain the unit tangent vector field representing the principal axis direction of the fast spatial target. and the unit normal vector field orthogonal to it , respectively represented as: , , in, This represents a small constant that prevents division by zero, for example... , This represents the x-direction component of the unit tangent vector. This represents the y-direction component of the unit tangent vector.

[0031] After sub-step 1, the length and orientation of the "stripes" where the fast target in space is located are predicted, and sub-step 2 is performed based on the prediction results.

[0032] Sub-step 2: Dynamically generate a rectangular sampling grid at each pixel position based on the length field and direction field. The rectangular sampling grid is stretched along the tangent vector direction (length is determined by...). Decide, The effective convolution kernel length is represented by a value that maintains a specific width along the normal vector direction. The actual position is calculated using the sampling points corresponding to the size of the effective convolution kernel.

[0033] Based on the normalized local relative coordinates, length field, direction field corresponding to the preset sampling point index, and the convolution kernel width corresponding to the current width branch, the actual sampling position of each sampling point in the global coordinate system of the image is calculated; a rectangular sampling grid is dynamically generated according to the actual sampling position, and the rectangular sampling grid is stretched along the unit tangent vector direction and maintains the width set by the corresponding width branch along the unit normal vector direction.

[0034] The embodiments of the present invention use The index of the sampling point corresponding to the effective convolution kernel size is represented by... Let the center pixel coordinates of the effective convolution kernel be represented, then the... The actual position of each sampling point in the global coordinate system of the image. The formula for calculation is: ; in, Represents normalized local relative coordinates, , Indicates the effective kernel length. This indicates the kernel width set for the current branch. (Based on the actual position...) The calculation formula enables the sampling points to be adaptively rotated and stretched according to the shape of the target in space.

[0035] Sub-step 3: Perform bilinear interpolation on the actual location to generate a resampled feature map.

[0036] Due to the actual position obtained by calculation Typically, these are floating-point numbers, meaning they are not located at integer pixel grid points. This invention addresses this problem using bilinear interpolation. A bilinear interpolation operator is employed. The pixel values ​​at corresponding positions are obtained from the input feature map to generate a resampled feature map, which is referred to as a resampled feature map in this embodiment of the invention. This indicates that the expression for the resampled feature map is: , Geometrically, sub-step 3 is equivalent to "correcting" the spatial fast target features of arbitrary orientation and length in the original spatial fast target image to be identified into features of standard orientation.

[0037] Sub-step 4: Preset a set of widths, and perform sub-step 2 and sub-step 3 on each width in the width set to obtain multiple resampled feature maps. Perform convolution operation on the resampled feature maps to obtain branch features. Perform weighted summation on the branch features to obtain the final features.

[0038] Since fast-moving targets in space may belong to different celestial bodies, the widths of the stripes from different targets are not necessarily the same. For example, the stripes produced by satellites with different orbital altitudes will not have the same width. In this embodiment of the invention, the orientation-adaptive rectangular convolution module is configured with multiple parallel branches of different widths. Specifically, a set of widths is preset. , For each width in the width set, sub-steps 2 and 3 are performed separately to obtain sampled features of different widths, thus obtaining multiple resampled feature maps. Convolution operations are then performed on the resampled feature maps to obtain branch features. .

[0039] A gated subnetwork is used to predict pixel-level weight maps corresponding to each width branch based on the input feature map or the features of each branch, and the pixel-level weight maps are normalized using the Softmax function; subsequently, the normalized pixel-level weight maps are... The features are applied to their respective branches, and a weighted sum is performed on each branch to obtain the final feature. final features Represented as: ; in, This represents element-wise multiplication. Indicates a branch of width.

[0040] Sub-step 5: Apply an affine transformation to the final features to obtain the encoder's output features. .

[0041] To introduce spatial adaptability, embodiments of the present invention address the final features. Applying affine transformation, utilizing subnetworks Hezi Network Predicting the matrix of affine transformation sum matrix The output features of the encoder are obtained. , is represented as: , in, This represents element-wise multiplication. This indicates element-wise addition.

[0042] The encoder's output features are fed into the decoder. In this embodiment, the decoder is connected to the MSHNet multi-scale prediction head. After the output features enter the decoder, they are first upsampled. The decoder upsamples the deep features level by level and concatenates them with the features of the corresponding levels of the encoder through skip connections. Prediction branches are derived at different levels of the decoder (e.g., layers 1 to 4). Each prediction branch compresses the feature channels to 1 through a 1×1 convolutional layer and is activated by a sigmoid activation function to generate prediction maps of different resolutions. Finally, the prediction maps of all scales are upsampled to the same size as the input feature maps and fused. Finally, a fusion convolutional layer is used to output the final stripe probability map. .

[0043] This invention abandons conventional cross-entropy loss or simple IoU loss, and improves the loss function of the model used to detect fast spatial targets. This invention constructs a hybrid loss function of physical perception and skeleton anchoring (PISA-Loss) as the total loss function of the improved feature extraction network, to overcome the difficulties of "elongated, blurry, and highly imbalanced" samples in fast spatial targets during the training phase. The expression for the total loss function is: ; in, This represents the skeleton anchoring weighted loss. This represents the loss of geometric consistency during decoupling. This represents the dynamic weighting coefficient.

[0044] The calculation of the total loss function includes the following three parts: Part 1: Calculating the Weighted Loss of Skeleton Anchoring .

[0045] The calculation of the skeleton anchoring weighted loss aims to make the improved feature extraction network prioritize the central skeleton region of the spatially fast target and reduce the impact of label uncertainty in blurred edge regions on the training process. Specifically, the central skeleton of the spatially fast target is first extracted based on the ground truth mask, and the Euclidean distance from each pixel to the central skeleton is calculated. Secondly, a spatial weight matrix is ​​constructed based on Euclidean distance. This approach assigns higher supervision weights to pixels closer to the central skeleton and relatively lower supervision weights to pixels farther from the central skeleton and located in blurred edge regions. Finally, the spatial weight matrix is ​​incorporated into the focus loss to obtain the skeleton anchoring weighted loss. .

[0046] The expression for the weight matrix is: , in, This represents the basic weight (the minimum contribution to retain the edge). The Gaussian kernel parameter represents the concentration of the control weights.

[0047] Finally, the weight matrix is ​​incorporated into the Focal Loss formula to calculate the skeleton anchoring weighted loss. : ; in, Represents the normalization factor. This represents the two-dimensional coordinates of the corresponding pixel in the prediction probability map, the ground truth mask, and the spatial weight map. Indicates the predicted probability. This indicates the focus parameter.

[0048] Part Two: Calculating the Decoupling Geometric Consistency Loss .

[0049] The core objective of the second part of the calculation is to force the predicted spatial fringe morphology to be consistent with the actual spatial fringe morphology through physical moment features, preventing problems such as breakage, bending, adhesion, or degradation into "spots" in the prediction results. Since spatial fringes are inherently linear, conventional pixel-level loss is insufficient to effectively penalize topological errors such as breakage, bending, or adhesion. This invention introduces set constraints based on second-order central moments, specifically including: First, moment feature extraction is performed. Then, the predicted probability maps are calculated. Truth Plot The normalized first-order moment and second-order central moment, i.e., the centroid Covariance Matrix Secondly, regarding the covariance matrix... Perform eigenvalue decomposition to obtain eigenvalues. and and principal feature vectors .in, and Let the squares of the principal axis length and the squares of the secondary axis length of the target be represented respectively. This indicates the principal direction of the spatial fringes. Finally, physical constraints are calculated, including a centroid positioning term, a direction alignment term, and a linearity (aspect ratio) constraint term. The centroid positioning term is expressed as... Used to constrain center position offset; direction alignment term is represented as The linearity (aspect ratio) constraint term is expressed as: This is used to ensure that predicted fast spatial targets appear as thin stripes rather than circular spots. Finally, the decoupling geometric consistency loss is calculated. Represented as: , in, This represents the centroid of the prediction probability graph. Denotes the centroid of the truth graph. This represents a small constant used to prevent division by zero. Indicates the weighting coefficient. This indicates the weight of the corresponding item. This represents the first principal eigenvector of the covariance matrix corresponding to the predicted probability map. This represents the first principal eigenvector of the covariance matrix corresponding to the truth map. This represents the first eigenvalue of the covariance matrix corresponding to the predicted probability map. This represents the second eigenvalue of the covariance matrix corresponding to the predicted probability map. This represents the first eigenvalue of the covariance matrix corresponding to the truth map. This represents the second eigenvalue of the covariance matrix corresponding to the truth map. The centroid positioning term is used to constrain the deviation of the predicted space target's center position from the actual space target. This represents the principal axis consistency term, used to constrain the consistency of the principal axis directions between the predicted space target and the actual space target. The linearity term is used to constrain the aspect ratio of the predicted spatial target to that of the real spatial target, so that the prediction result maintains a thin stripe-like structure. Part Three: Adjusting Dynamic Course Learning Using Dynamic Weighting Coefficients.

[0050] To prevent gradient oscillations caused by strong geometric constraints in the early stages of training, this embodiment of the invention designs dynamic weight coefficients. The dynamic weighting coefficient is related to the average intersection-union ratio (Batch IoU) of the current batch. Represented as: , in, This represents the average intersection-over-union ratio (IoU) between the predicted probability map and the ground truth mask in the current training batch, used to determine whether the model has entered a relatively stable learning phase. The hyperparameters that control the steepness of the Sigmoid curve are... This indicates the preset intersection-union ratio threshold, used to determine whether the model has entered a relatively stable learning phase.

[0051] In the early stages of training the improved feature extraction network, the intersection-union ratio (IU) is... The dynamic weight coefficients are relatively low, approaching 0. The learning of the improved feature extraction network is mainly guided by the skeleton anchoring weighted loss, enabling it to learn pixel classification. As the improved feature extraction network converges, the intersection-union ratio (IU / U) increases. As the dynamic weight coefficient increases, a decoupling geometric consistency loss is introduced to refine the constraint on the spatial stripe morphology.

[0052] S2: Training an improved feature extraction network.

[0053] Acquire training samples with a spatial fast ground truth mask; input the training samples into an improved feature extraction network; output a stripe probability map through an encoder, OARConv module, decoder, and multi-scale prediction head; calculate L_total based on the predicted probability map and the ground truth mask; update the network parameters using backpropagation, iterating until the preset convergence condition is met.

[0054] S3: Utilize the improved feature extraction network after training to detect spatial targets quickly.

[0055] The improved feature extraction network, after training, directly outputs a multi-scale fused stripe probability map during the inference phase. By setting a threshold to binarize the stripe probability map, a precise segmentation mask for fast-moving targets in space can be obtained, thereby achieving the detection and localization of fast-moving targets in space.

[0056] Although embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

[0057] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for rapid target detection in space, characterized in that, include: S1: Improve the traditional U-Net architecture to obtain an improved U-Net architecture, and use the improved U-Net architecture as the backbone network to construct an improved feature extraction network; The improvement method of the U-Net improved architecture is as follows: in the encoder part of the traditional U-Net architecture, the standard convolutional layer in the basic building unit is replaced with an orientation-adaptive rectangular convolutional module. The total loss function of the improved feature extraction network is: ； in, This represents the skeleton anchoring weighted loss. This represents the loss of geometric consistency during decoupling. Indicates the dynamic weighting coefficient; S2: Train the improved feature extraction network; S3: Detect spatial targets quickly using the improved feature extraction network after training.

2. The rapid space target detection method as described in claim 1, characterized in that, The orientation-adaptive rectangular convolution module includes a length prediction subnetwork and an orientation prediction subnetwork, which are used to calculate the length field and orientation field of the prediction convolution kernel from the input feature map, respectively.

3. The space rapid target detection method as described in claim 2, characterized in that, The orientation-adaptive rectangular convolution module includes the following sub-steps: Sub-step 1: Utilize the length prediction sub-network and the orientation prediction sub-network to solve the input feature map pixel by pixel, and predict the length field and orientation field of the effective convolutional kernel at each pixel; Sub-step 2: Based on the length field and direction field, dynamically generate a rectangular sampling grid at each pixel position. The rectangular sampling grid is stretched along the tangent vector direction and maintains a specific width along the normal vector direction. Calculate the actual position using the sampling points corresponding to the size of the effective convolution kernel. Sub-step 3: Perform bilinear interpolation on the actual location to generate a resampled feature map; Sub-step 4: Preset a set of widths, and perform sub-step 2 and sub-step 3 on each width in the width set to obtain multiple resampled feature maps; Perform a convolution operation on the resampled feature map to obtain branch features; The final feature is obtained by weighted summation of the branch features.

4. The rapid space target detection method as described in claim 3, characterized in that, The sub-steps of the orientation-adaptive rectangular convolution module also include: Sub-step 5: Apply an affine transformation to the final features to obtain the encoder's output features. , is represented as: ； in, This indicates that the final feature is obtained by weighted summation of the branch features. This represents element-wise multiplication. This represents element-wise addition. and Both represent matrices for predicting affine transformations.

5. The rapid space target detection method as described in claim 3, characterized in that, Each pixel The length of the effective convolution kernel at the location Represented as: ； in, Indicates the length scaling factor. This represents the bias factor. This represents a single-channel scalar graph. To predict from the input feature map using the length prediction subnetwork; The orientation prediction subnetwork is used to predict a two-channel vector map from the input feature map. The two-channel vector map is then L2 normalized to obtain a unit tangent vector field representing the principal axis direction of a fast spatial target. and the unit normal vector field orthogonal to it , is represented as: ； ； in, Represents a two-channel vector graph. This represents a small constant used to prevent division by zero. This represents the x-direction component of the unit tangent vector. This represents the y-direction component of the unit tangent vector.

6. The rapid space target detection method as described in claim 3, characterized in that, The method for calculating the actual position is as follows: ； in, This represents the center pixel coordinates of the effective convolution kernel. This represents the index of the sampling point corresponding to the effective convolution kernel size. Represents normalized local relative coordinates. , Indicates the effective convolution kernel length. This indicates the kernel width set for the current branch.

7. The space rapid target detection method as described in claim 3, characterized in that, The resampled feature map Represented as: ； in, Indicates the actual location. Indicates the input feature map, This represents the bilinear interpolation operator.

8. The rapid space target detection method as described in claim 3, characterized in that, The final feature Represented as: ； in, Indicates branching characteristics, This represents the pixel-level weight map predicted by the gated subnetwork. Indicates a branch of width.

9. The space-based rapid target detection method as described in claim 4, characterized in that, After obtaining the output features, the output features are fed into the decoder. The decoder is connected to the MSHNet multi-scale prediction head. The multi-scale prediction head generates prediction maps at different resolution levels of the decoder, and then upsamples the prediction maps to the size of the input feature map and fuses them to obtain a stripe probability map.

10. The method for rapid target detection in space as described in claim 1, characterized in that, The calculation process of the total loss function consists of three parts: Part 1: Calculating the Weighted Loss of Skeleton Anchoring : ； in, Represents the normalization factor. This represents the two-dimensional coordinates of the corresponding pixel in the prediction probability map, the ground truth mask, and the spatial weight map. Indicates the predicted probability. Indicates the focus parameter, Represents a spatial weighted graph; Part Two: Calculating the Decoupling Geometric Consistency Loss : ， in, This represents the centroid of the prediction probability plot. Denotes the centroid of the truth graph. This represents a small constant used to prevent division by zero. Indicates the weighting coefficient. Indicates the weight of the corresponding item. This represents the first principal eigenvector of the covariance matrix corresponding to the predicted probability map. This represents the first principal eigenvector of the covariance matrix corresponding to the truth map. This represents the first eigenvalue of the covariance matrix corresponding to the predicted probability map. This represents the second eigenvalue of the covariance matrix corresponding to the predicted probability map. This represents the first eigenvalue of the covariance matrix corresponding to the truth map. This represents the second eigenvalue of the covariance matrix corresponding to the truth map. Indicates the centroid positioning term. Indicates the consistency term in the main direction. Represents the linearity term; Part Three: Adjusting Dynamic Course Learning Using Dynamic Weighting Coefficients, whereby the dynamic weighting coefficients are expressed as follows: ； in, This represents the average intersection-union ratio (IoU) between the predicted probability map and the ground truth mask in the current training batch. The hyperparameters that control the steepness of the Sigmoid curve are... This indicates the preset crossover ratio threshold.

Images