Space stripe target detection method based on physical prior driven self-supervised network

Abstract

The present application belongs to the field of space target detection, and particularly relates to a space stripe target detection method based on a physical prior driven self-supervised network. A physical prior encoding module in the network is used for explicitly encoding the physical imaging prior of the space stripe target to generate a physical response map. A multi-scale feature extraction module in the network is used for extracting a multi-scale feature map from the physical response map. A physical consistency attention module is used for querying a main direction and a width from the physical response map, calculating a transverse point spread function consistency and a longitudinal energy continuity score, generating an attention map and residual modulating each scale feature. A stripe direction perception lightweight detection module in the network performs direction perception feature enhancement through a multi-direction re-parameterized strip convolution block, and adopts a high-resolution multi-scale fusion structure for multi-scale feature fusion. A detection head in the network is used for outputting a detection frame. The present application can effectively improve the stripe detection precision, reduce the false alarm rate, and meet the lightweight deployment requirements.

Description

Technical Field

[0001] This invention belongs to the field of space target detection technology, and in particular relates to a spatial stripe target detection method based on a physical prior-driven self-supervised network. Background Technology

[0002] Space fringe target detection is a crucial technical component in space situational awareness, space debris monitoring, low-Earth orbit satellite tracking, astronomical observation image quality control, and ground-based / spaceborne optical surveillance systems. During long-exposure optical imaging, high-speed moving space targets undergo continuous displacement relative to the stellar background. Their energy integrates along their trajectory over the exposure time, ultimately forming linear fringes with distinct directionality and a large aspect ratio on the image. Unlike conventional point targets, area targets, or natural image targets, space fringe targets typically exhibit extremely slender structures with a width of only a few pixels and a length of tens to hundreds of pixels. Their orientation can cover any angle within the image plane, and their brightness distribution is influenced by multiple factors, including target reflectivity, exposure time, relative angular velocity, optical system point spread function, atmospheric disturbance, readout noise, and the background star field. Under low signal-to-noise ratio conditions, the boundaries of fringe targets are often indistinct, and local areas may exhibit breaks, energy attenuation, or be submerged by the stellar background, making automated detection tasks highly challenging.

[0003] From the perspective of optical imaging mechanisms, spatial fringe targets are not ordinary line segments or geometric shapes, but rather the result of the combined effect of their motion trajectory and the point spread function of the imaging system. Ideally, a high-speed target can be approximated as moving at a constant speed along a straight line during exposure. Its imaging result can be understood as the convolution of a line segment function extending along the direction of motion with the optical system's point spread function, followed by the addition of noise. Therefore, real spatial fringes not only possess continuity along their major axis but also exhibit finite broadening laterally due to the point spread function. This physical characteristic fundamentally distinguishes them from bright linear interferences such as cosmic ray hits, hot pixel strings, bad columns, diffraction spikes, and saturated star tails. Cosmic ray hits are typically generated by high-energy particles directly impacting the detector, without optical imaging, and lack typical point spread function broadening in their lateral profile. Hot pixel strings or bad columns are often distributed along fixed rows and columns, with discontinuous intensity variations. Diffraction spikes are usually associated with the center of bright stars, exhibiting a clear central concentration of energy. These interferences are visually similar to real spatial fringes, but their physical generation mechanisms differ, making them a significant source of false alarms.

[0004] Existing technologies for spatial stripe target detection have several shortcomings, making it difficult to obtain stable, reliable, and efficient stripe detection results, as follows: First, traditional geometric transformation and template matching methods are highly dependent on manual parameters and scene conditions, resulting in insufficient robustness. Existing methods based on Radon transform, Hough transform, orientation template matching, threshold segmentation, morphological processing, or multi-stage hypothesis testing typically require pre-setting detection thresholds, filtering templates, orientation sampling intervals, cumulative voting rules, or candidate selection conditions. These methods can achieve certain detection results when the target is bright, the background is simple, and the noise is low. However, in real-world spatial observation images, striped targets are often in environments with low signal-to-noise ratios, strong star backgrounds, stray light interference, and noise disturbances, and may exhibit breaks, local energy attenuation, or overlap with background stars. Due to the lack of end-to-end learning capabilities and adaptive feature representation capabilities in traditional methods, their parameters often struggle to remain stable across different telescopes, exposure conditions, and background complexities, easily leading to missed detections, false detections, and increased false alarm rates.

[0005] Second, existing supervised deep learning methods rely on large amounts of real, manually labeled data, making them ill-suited for application scenarios where real spatial stripe samples are scarce. Real-world samples of spatial stripes are typically difficult to obtain, and stripes constitute a very small percentage of images. Manual annotation requires determining the start and end positions, boundary ranges, or detection box locations of thin stripes, resulting in high annotation costs, long processing times, and significant subjective variations. For weak, blurred, and broken stripes, different annotators may have differing judgments about boundaries and endpoints, leading to insufficient annotation consistency. Existing deep learning detectors typically require a large number of accurately labeled samples to achieve good generalization ability. Once the training data scale is insufficient or real observation conditions change, the model is prone to overfitting, performance degradation across scenes, and insufficient adaptability to unknown backgrounds.

[0006] Third, existing general-purpose convolutional neural networks lack an explicit expression of the physical imaging mechanism of spatial stripes. Spatial stripe targets are essentially the imaging result of the convolution of moving line segments with the point spread function of the optical system, superimposed with noise. Their lateral profile, longitudinal extension direction, and energy distribution all have clear physical meanings. However, existing depth detection methods mostly use ordinary convolutional layers to learn features from a randomly initialized state. The morphology of the convolutional kernel usually lacks physical constraints corresponding to the stripe imaging model, making it difficult to directly obtain stripe response capabilities similar to matched filters at the network front end. This leads to the network needing to rely on a large number of samples to learn the basic morphology of weak, thin stripes. When real annotations are scarce or training samples are limited, problems such as low feature learning efficiency, insufficient response to weak stripes, and unstable detection results easily arise.

[0007] Fourth, existing detection models struggle to effectively distinguish between genuine motion stripes and bright linear interference. In real starscape images, besides genuine spatial stripe targets, there may also be bright linear interference such as cosmic ray strikes, hot pixel clusters, bad columns, diffraction spikes, and saturated star trails. These interferences may also appear visually as localized lines, highlights, elongated shapes, or directional structures, exhibiting high similarity to genuine motion stripes. Existing deep learning models typically rely primarily on appearance texture and statistical data features for discrimination, lacking explicit constraints on physical consistency characteristics such as "genuine stripes should have lateral point spread function broadening and longitudinal energy continuous extension." Consequently, in complex backgrounds, they easily misclassify false stripes as genuine targets, leading to an increased false alarm rate.

[0008] Fifth, existing attention mechanisms are mostly data-driven, lacking physical interpretability and targeted interference suppression capabilities. General visual attention mechanisms typically enhance salient region features through channel weights, spatial weights, or self-attention calculations, but their attention maps are primarily generated by training data, lacking interpretable criteria corresponding to the physical imaging process of spatial stripes. Under low signal-to-noise ratio conditions, purely data-driven attention may simultaneously enhance responses to real stripes and interference such as local noise, cosmic rays, hot pixel strings, or diffraction spikes, making it difficult to form a stable false alarm suppression mechanism. Therefore, while existing attention methods can improve feature representation capabilities, they struggle to specifically address the problem of "difficulty in distinguishing between bright linear interference and real stripes" in spatial stripe detection.

[0009] Sixth, existing general-purpose multi-scale detection structures are prone to losing weak, thin stripe information during deep downsampling. Spatial stripe targets are typically only a few pixels wide, but their length is much greater than their width, exhibiting extremely thin and elongated features and high-frequency edge characteristics. General-purpose target detection networks usually employ multi-level downsampling and deep semantic feature fusion, which is beneficial for the semantic recognition of ordinary targets. However, for weak stripe targets with extremely narrow widths, after downsampling at 16x, 32x, or even higher magnifications, the lateral stripe structure may degenerate into a near-single-pixel response, or even be completely submerged by convolution and smoothing operations. If existing general-purpose feature pyramid structures are directly used for spatial stripe detection, they are prone to causing weak stripe breaks, missed detection of small targets, and decreased localization accuracy.

[0010] Seventh, existing high-precision detection networks have a large computational and parameter load, making it difficult to meet the lightweight deployment requirements of spaceborne or ground-based edge platforms. Space target monitoring missions typically require processing large batches of continuous observation images, and some application scenarios also require rapid detection on spaceborne equipment, small ground-based telescope edge computing platforms, or real-time early warning systems. To improve accuracy, existing deep detection models often employ deep backbone networks, complex attention structures, or computationally intensive feature fusion modules, resulting in significant inference latency and storage overhead, which is not conducive to deployment on platforms with limited computing power, power consumption, and storage resources. For large-scale sky survey images or real-time space situational awareness systems, insufficient detection efficiency directly affects data processing throughput and the timeliness of target detection.

[0011] Eighth, existing training methods based on synthetic data or pseudo-labels are complex and prone to introducing multi-source pseudo-label errors. To alleviate the problem of insufficient real annotations, some methods use synthetic stripes or pseudo-labels to construct training data. However, this often requires generating multiple labels simultaneously, such as orientation, length, width, mask, boundary, centerline, and energy distribution, and using multiple auxiliary loss functions for joint training. While these methods can increase training supervision information, they also increase the complexity of the training process and introduce the problem of balancing multi-task loss weights. When some pseudo-labels are not completely consistent with the real imaging physics, the error may also be propagated to the network optimization process, affecting model stability and engineering reproducibility.

[0012] Ninth, existing methods lack interpretability, making it difficult to analyze and verify detection results at the level of physical parameters. Space target optical observation has a clear imaging physics background. If the detection model can output or implicitly reflect physical quantities such as fringe direction, length, and dot spread function width, it will facilitate result analysis, system debugging, and reliable engineering applications. However, existing general-purpose depth detectors typically only output detection boxes, class probabilities, or segmentation masks. The internal features lack clear physical meaning, making it difficult to explain why the network enhances certain fringe responses or suppresses certain interference responses, and also difficult to analyze in relation to the telescope's dot spread function, exposure conditions, and target motion characteristics. Summary of the Invention

[0013] In view of this, the present invention aims to provide a spatial stripe target detection method based on a physical prior-driven self-supervised network to solve the problems of insufficient detection reliability of existing technologies under conditions such as scarce real annotations, complex backgrounds with low signal-to-noise ratio, dense bright linear interference, easy loss of weak and thin stripes, and limited computing power of edge platforms.

[0014] To achieve the above objectives, the technical solution created by this invention is implemented as follows: A spatial stripe target detection method based on a physics prior-driven self-supervised network includes the following steps: S1: Establish a self-supervised fringe detection network based on physical priors, which includes a physical prior encoding module, a multi-scale feature extraction module, a physical consistency attention module, a fringe direction perception lightweight detection module, and a detection head; S2: Acquire spatial observation images and input them into the physical prior coding module. Based on the physical parameterized convolutional kernel group, the physical imaging prior of the spatial stripe target is explicitly encoded to generate a physical response map. S3: Input the physical response map into the multi-scale feature extraction module to extract at least three multi-scale feature maps of different scales; S4: Input the multi-scale feature map into the physical consistency attention module, determine the estimated principal direction and the estimated point spread function width based on the physical response map and the physical parameter set corresponding to the response in each direction, calculate the horizontal point spread function consistency score and the vertical energy continuity score respectively, fuse them to generate the physical consistency attention map, and modulate the multi-scale feature map; S5: Input the attention-modulated multi-scale feature map into the stripe direction-aware lightweight detection module, enhance the direction-aware features through multi-directional reparameterized stripe convolutional blocks, and perform multi-scale feature fusion using a high-resolution multi-scale fusion structure to output the enhanced multi-scale feature map; wherein, the fusion weight is jointly determined by the learnable fusion weight and the physical response gating after global average pooling of the physical response map; S6: Input the enhanced multi-scale feature map into the detection head to predict the spatial stripe target detection box.

[0015] Furthermore, the physics prior coding module defines a set of physical parameters consisting of M directional anchor points: ; in, This represents the direction angle of the m-th anchor point. This represents the length of the m-th stripe template. This represents the width of the point spread function of the m-th stripe template; Generate a set of physically parameterized convolutional kernels based on the set of physical parameters.

[0016] Furthermore, the generation method for the m-th stripe template in the physically parameterized convolution kernel group is as follows:

[0017] ; ; in, This represents the fringe template response value at coordinates (x, y) in the spatial observation image; Represents the normalization constant; Represents the coordinates along the principal direction of the stripes; Represents the coordinates perpendicular to the principal direction of the stripes; Using the Sigmoid function for rectangle functions Smoothing is performed to make the stripe template aligned with the direction angle. Stripe template length and point spread function width Preserving differentiability:

[0018] in, This represents the Sigmoid activation function. This represents the hyperparameter used to control the width of the transition band.

[0019] Furthermore, the physical response diagram is generated as follows: For an input spatial observation image I, the orientation response map generated by the m-th fringe template for: ; in, Represents a two-dimensional convolution operation; The M directional response maps are concatenated along the channel dimension and then mixed using a 1×1 convolution to obtain the physical response map: ; in, This represents the 1×1 convolution weight.

[0020] Furthermore, the multi-scale feature extraction module extracts a first-scale feature map with a downsampling factor of 4, a second-scale feature map with a downsampling factor of 8, and a third-scale feature map with a downsampling factor of 16 from the physical response map. The first-scale feature map, the second-scale feature map, and the third-scale feature map together form a multi-scale feature map.

[0021] Furthermore, the physical consistency attention module determines the estimated principal direction based on the physical response map and the physical parameter set corresponding to the responses in each direction. And the estimated point spread function width is ; The consistency score of the lateral point spread function is calculated as follows: Centered on coordinates (x, y), along a direction perpendicular to the estimated principal direction In the direction, a length of [value] is extracted from the multi-scale feature map through bilinear sampling. cross section : ; in, Indicates along the direction perpendicular to the estimated principal direction The offset; Indicates the sampling radius of the transverse profile; Represents the feature map at the s-th scale; Construction and width of point spread function Corresponding Gaussian template :

[0022] The lateral profile is normalized and matched with the Gaussian template to obtain the consistency score of the lateral point spread function. :

[0023] in, and These represent the mean-removed cross-section and the Gaussian template, respectively. The longitudinal energy continuity fraction is calculated as follows: Centered on coordinates (x, y), along the estimated principal direction Bilinear sampling is used to extract a length of [length missing] from the multi-scale feature map. longitudinal section : ; in, Indicates the sampling radius of the longitudinal profile; According to the longitudinal section Constructing longitudinal energy continuity fractions : .

[0024] Furthermore, the horizontal point spread function consistency score and the vertical energy continuity score are weighted and fused to generate a physically consistent attention map: ; in, They represent learnable scalars, This represents the Sigmoid activation function; The physically consistent attention map modulates the multi-scale feature map using a residual weighting method: ; in, This represents the modulated multi-scale feature map; Represents the multi-scale feature map before modulation; This indicates element-wise multiplication.

[0025] Furthermore, the multi-directional reparameterized strip convolutional block adopts a multi-branch structure during the training phase, including a k×k depth-separable convolutional branch, a 1×k horizontal strip convolutional branch, a k×1 vertical strip convolutional branch, a 1×1 point convolutional branch, and an identity mapping branch. Among them, the k×k depth-separable convolutional branch is used to extract local features, the 1×k horizontal strip convolutional branch is used to enhance the horizontal stripe response, the k×1 vertical strip convolutional branch is used to enhance the vertical stripe response, the 1×1 point convolutional branch is used for channel fusion, and the identity mapping branch is used to keep the residual information unchanged. Output feature map of multi-directional reparameterized strip convolutional blocks during the training phase for:

[0026] in, This represents the output of the i-th convolutional branch. When i=1, it is the output of the k×k depth separable convolutional branch; when i=2, it is the output of the 1×k horizontal strip convolutional branch; when i=3, it is the output of the k×1 vertical strip convolutional branch; and when i=4, it is the output of the 1×1 point convolutional branch. Indicates the first Batch normalization operation for each convolutional branch This represents the input feature map of the multi-directional reparameterized strip convolutional block at the current scale. During the usage phase, the multi-branch structure from the training phase is folded into a single equivalent convolution through structure reparameterization; for any combination of convolution and batch normalization, the equivalent weights and biases are:

[0027]

[0028] in, Represents the convolution weights. These represent the batch normalization parameters, A constant that prevents division by zero; The kernels of the 1×k horizontal strip convolution branch, the k×1 vertical strip convolution branch, and the 1×1 point convolution branch are expanded to k×k size using zero padding. The identity mapping branch is represented as a unit impulse convolution kernel. Finally, the equivalent weights and biases of all branches are summed to obtain a single-path convolution. ; in, This represents the output feature map during the usage phase; This represents the equivalent convolution kernel after fusion; Indicates the convolution operation; This indicates the bias after fusion.

[0029] Furthermore, the high-resolution multi-scale fusion structure adopts a lightweight bidirectional feature pyramid structure. The fusion weights of the lightweight bidirectional feature pyramid structure are determined as follows:

[0030]

[0031] in, This represents the enhanced multi-scale fused feature map; This represents the feature map at the i-th scale input to the lightweight bidirectional feature pyramid structure; This represents the fusion weight at the i-th scale; Represents the learnable fusion weights at the i-th scale. This represents the physical response map downsampled to the i-th scale. Indicates a learnable scalar. This indicates global average pooling.

[0032] Furthermore, during the training phase of the self-supervised fringe detection network, training samples are constructed using a single-label physical injection method, specifically including: First, input an unlabeled, realistic background image. In real background image Local intensity maxima are detected as candidate stars, and a two-dimensional Gaussian fit is performed on the neighborhood of the candidate stars to estimate the width of the spread function at the image points. Simultaneously, robust noise standard deviation is calculated in the dark background region to obtain noise estimates. .

[0033] Secondly, in terms of the width of the image point spread function and noise estimates Sampling stripe parameters in the physical distribution centered on:

[0034] in, Indicates the direction of the stripes in this sampling; This indicates the stripe length in this sample; This represents the width of the point spread function in this sampling. This represents the radiant energy accumulated per unit length of the stripe along its trajectory. Indicates the starting point of the stripes; Based on the sampled stripe parameters Generate stripe signal ; Secondly, based on stripe signals Training images were generated using a Poisson electron counting model. :

[0035] in, This represents the Poisson electron counting model; Finally, align the circumscribed rectangle with the calculation axes based on the fringe start and end points, and expand it outwards. Boundary margin, generating unique pseudo-labels :

[0036] in, Indicates the center coordinates of the false label detection box. This indicates the width and height of the pseudo-label detection box.

[0037] Furthermore, the set of spatial stripe target detection boxes: ; in, This indicates the center coordinates of the spatial stripe target detection box. This represents the width and height of the spatial stripe target detection box. This indicates the confidence level of the spatial fringe target.

[0038] A computer device, comprising: At least one processor; and A memory that is communicatively connected to at least one processor; wherein, The memory stores instructions that can be executed by at least one processor, which in turn executes the above-described spatial stripe target detection method.

[0039] A non-transitory computer-readable storage medium storing computer instructions for causing a computer to execute the above-described spatial stripe target detection method.

[0040] Compared with the prior art, the present invention can achieve the following beneficial effects: (1) Reduce reliance on real manually labeled data and lower data construction costs. This invention uses unlabeled real background images as a basis, and automatically samples physically reasonable stripe parameters and generates synthetic training images by estimating the point spread function width and noise level. Training can be completed by simultaneously generating false labels for detection boxes. This method significantly reduces the need for manual annotation, lowers the cost of training data construction, and improves the applicability of the method in scenarios where real annotations are insufficient.

[0041] (2) Enhanced detection capability of weak and thin stripes, improving response stability under low signal-to-noise ratio conditions. This invention transforms the physical imaging law of observed stripes formed by the convolution of moving line segments and point spread functions into a learnable set of physically parameterized convolution kernels through a physical prior coding module. This enables the detection network front-end to directly possess an initial response capability consistent with the stripe imaging mechanism. Therefore, this invention can improve the feature response intensity and training efficiency of weak stripes, reducing the problem of inefficient learning of the basic stripe morphology from data using ordinary convolution.

[0042] (3) Improve the ability to suppress bright linear interference and reduce the false alarm rate. The physical consistency attention module provided by this invention explicitly calculates the lateral point spread function consistency and longitudinal energy continuity based on the estimated principal direction and estimated point spread function width provided by the physical prior coding module. Real motion stripes usually satisfy both lateral optical broadening and longitudinal continuous extension, while many types of false stripe interference do not possess complete physical consistency. Therefore, this invention can enhance the real stripe response and suppress the false stripe response from the physical generation mechanism, thereby reducing the false alarm rate in complex backgrounds.

[0043] (4) Better protection of high-resolution stripe information, improving detection integrity and positioning reliability. The stripe direction sensing lightweight detection module provided by this invention adopts a multi-scale fusion strategy biased towards high resolution, using only 4x downsampled features, 8x downsampled features, and 16x downsampled features. At the same time, it combines learnable fusion weights with the physical response gating of the physical prior coding module, so that scale features with stronger physical responses receive higher weights. This design can better preserve the detailed structure of weak stripes, reducing missed detections, breaks, and positioning errors.

[0044] (5) Balancing detection accuracy and inference efficiency for lightweight applications. The fringe direction sensing lightweight detection module provided by this invention adopts a reparameterized design with multiple branches in the training phase and a single branch in the inference phase. In the training phase, the fringe direction representation capability is enhanced through k×k convolution, horizontal stripe convolution, vertical stripe convolution, point convolution, and identity mapping branches. In the inference phase, the multiple branches are folded into a single equivalent convolution. Thus, this invention can reduce the computational overhead in the inference phase while maintaining strong fringe direction sensing capability, thereby improving the efficiency of large-scale spatial observation image processing. Attached Figure Description

[0045] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments and descriptions of the invention are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 A schematic diagram of the overall architecture of the self-supervised stripe detection network described in the embodiments of the present invention; Figure 2A schematic diagram of the structure of the computer device described in the embodiment of the present invention.

[0046] The reference numerals in the accompanying drawings include: 12, computer equipment; 14, external device; 16, processing unit; 18, bus; 20, network adapter; 22, I / O interface; 24, display; 28, system memory; 30, RAM; 32, cache memory; 34, storage system; 40, program / utility; 42, program module. Detailed Implementation

[0047] 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 merely illustrative of the invention and do not constitute a limitation thereof.

[0048] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0049] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," 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.

[0050] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "assembly," "connection," and "joining" 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.

[0051] The invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0052] This invention proposes a spatial stripe target detection method based on a physics prior-driven self-supervised network. Targeting weak, thin stripes formed by high-speed spatial targets during long exposures, this method constructs a physics prior-driven self-supervised stripe detection network. This allows it to fully utilize the physical imaging priors, point spread function characteristics, directional continuity, and high-resolution geometric features of spatial stripe targets without relying on extensive manual annotation. This enables stable detection of low signal-to-noise ratio, weak, thin, and arbitrarily oriented spatial stripe targets. It offers advantages in reducing false alarm rates, improving detection integrity, enhancing physical interpretability, and meeting lightweight deployment requirements.

[0053] like Figure 1 As shown, the self-supervised fringe detection network consists of a physical prior encoding module, a multi-scale feature extraction module, a physical consistency attention module, a fringe direction perception lightweight detection module, and a detection head. The multi-scale feature extraction module and the physical consistency attention module form the backbone network. This invention places the physical prior encoding module at the front end of the detection network, the backbone network in the middle section, and the fringe direction perception lightweight detection module at the rear end, enabling the backbone network and the fringe direction perception lightweight detection module to share the physical imaging prior of the spatial fringe target explicitly encoded by the physical prior encoding module.

[0054] The forward processing procedure of the self-supervised fringe detection network is as follows: First, the acquired spatial observation images are input into the physical prior coding module. Based on the physical parameterized convolution kernel group, the physical imaging prior explicit coding of the spatial stripe target is used to generate a physical response map.

[0055] Secondly, after the physical response map is adapted for channels and resolution, it is input into the multi-scale feature extraction module to extract multi-scale feature maps of at least three different scales.

[0056] Then, the multi-scale feature map is input into the physical consistency attention module. Physical parameters such as the estimated principal direction and the estimated point spread function width are queried from the physical response map shared by the physical prior coding module. The horizontal point spread function consistency score and the vertical energy continuity score are calculated and fused to generate the physical consistency attention map. The physical consistency attention map is then used to perform residual modulation on the multi-scale feature map to generate the attention-modulated scale feature map.

[0057] Next, the attention-modulated scale feature map is input into the stripe direction-aware lightweight detection module. The direction-aware feature is enhanced by multi-directional reparameterized stripe convolutional blocks, and multi-scale feature fusion is performed by a high-resolution multi-scale fusion structure to output the enhanced multi-scale feature map. The fusion weights are jointly determined by the learnable fusion weights and the physical response gating after global average pooling of the physical response map.

[0058] Finally, the enhanced multi-scale feature map is input into the detection head to predict the spatial stripe target detection box.

[0059] The five functional modules of the self-supervised stripe detection network will be explained in detail below.

[0060] I. Physical Prior Coding Module The physical prior coding module is used to explicitly encode the physical imaging prior of the spatial stripe target at the network front end. Then, the physical imaging law of the observed stripes formed by the convolution of moving line segments and point spread functions is embedded into the network front end in the form of a learnable physical parameter set. The physical parameter set generates a physical parameterized convolution kernel set, so that the front end of the stripe detection network directly has an initial response capability consistent with the stripe imaging mechanism.

[0061] High-speed space targets move at approximately a uniform linear velocity relative to the imaging system during exposure. The imaging result can be represented as the convolution of a single-pixel wide motion line segment with the system's point spread function, followed by the superposition of observation noise. ; in, This represents the radiant energy accumulated per unit length of the stripe along its trajectory. Indicates the origin of the stripes and the end of the stripes A defined single-pixel wide line segment function; The point spread function represents the imaging system; Indicates additive observation noise; This represents a two-dimensional convolution operation.

[0062] The point spread function is approximated as an isotropic Gaussian function. Then the direction angle is The stripe template length is The width of the point spread function is The stripe template can be represented as:

[0063]

[0064]

[0065] in, Represents the normalization constant; Represents the coordinates along the principal direction of the stripes; Represents the coordinates perpendicular to the principal direction of the stripes.

[0066] To enable end-to-end backpropagation of template parameters during training, the Sigmoid function is used for the rectangular function. Perform smoothing:

[0067] in, This represents the Sigmoid activation function. This represents the hyperparameter used to control the width of the transition band. This smoothing process adjusts the stripe template for the direction angle. Stripe template length and point spread function width The parameters remain differentiable.

[0068] The physical prior coding module defines a set of physical parameters consisting of M directional anchor points: ; in, This represents the direction angle of the m-th anchor point. This represents the length of the m-th stripe template. This represents the width of the point spread function for the m-th stripe template. (Using...) and The logarithmic parameterization method ensures that the length and width parameters are positive values. During initialization, it can be... exist The stripe template is evenly distributed within the interval, so that it covers spatial stripes in any direction within the image plane.

[0069] For an input spatial observation image I, the orientation response map generated by the m-th fringe template for: ; The M directional response maps are concatenated along the channel dimension and then channel-blended using a 1×1 convolution to obtain the physical response map. : ; in, This represents the 1×1 convolution weight.

[0070] The physical response map serves as a shared physical information source for the detection network: on the one hand, after channel and resolution adaptation, it is used as the input feature of the backbone network; on the other hand, combined with the corresponding physical parameter set, it provides local principal direction and point spread function width information for the physical consistency attention module, and after scale transformation, it serves as the physical response gating basis in the multi-scale fusion of the stripe direction perception lightweight detection module.

[0071] II. Multi-scale Feature Extraction Module The physical response map output by the physical prior encoding module is first adapted to the number of channels and resolution through stem convolution, and then input into the multi-scale feature extraction module. This multi-scale feature extraction module can adopt any convolutional neural network structure suitable for single-stage object detection. Its main function is to extract spatial features of different scales from the physical response map to form a multi-scale feature map.

[0072] In one example of the present invention, feature maps of three different scales are extracted from the physical response map. :

[0073] in, This is the first-scale feature map with a sampling ratio of 4. This is the second-scale feature map with a downsampling factor of 8. This is a third-scale feature map with a downsampling factor of 16; The input is the physical response diagram; This is the backbone network transformation function, representing the nonlinear transformation operations such as convolution, pooling, and activation performed by the backbone network.

[0074] III. Physical Consistency Attention Module The Physically Consistent Attention module is used to construct physically consistent attention on multi-scale feature maps to suppress bright line-like interference such as cosmic rays, hot pixel strings, and diffraction spikes. The core idea of ​​the Physically Consistent Attention module is that real spatial fringes should simultaneously satisfy lateral point spread function consistency and longitudinal energy continuity, while some pseudo-fringe interferences, although visually appearing as bright lines, do not satisfy the above physical characteristics.

[0075] For feature maps of any scale , , Indicates the number of channels. Indicates the height of the feature map. Indicates the width of the feature map.

[0076] The physical consistency attention module determines the estimated principal direction based on the physical response map and the physical parameter set corresponding to the response in each direction. and the estimated point spread function width In one implementation, the direction angle corresponding to the direction anchor point with the largest local response value can be selected as the estimated principal direction. The point spread function width of the corresponding anchor point is used as the estimated point spread function width. Alternatively, statistical values ​​of the spread function width parameters at multiple learnable points can be used as the estimated spread function width. .

[0077] The Physical Consistency Attention module has two physical constraint functions: The first physical constraint is the consistency of the lateral point spread function, denoted as... Centered on coordinates (x, y), along a direction perpendicular to the estimated principal direction. In the direction, a length of [value] is extracted from the multi-scale feature map through bilinear sampling. cross section : ; in, Indicates along the direction perpendicular to the estimated principal direction The offset; Indicates the sampling radius of the transverse profile; Represents the feature map at the s-th scale; Construction and width of point spread function Corresponding Gaussian template :

[0078] The lateral profile is normalized and matched with the Gaussian template to obtain the consistency score of the lateral point spread function. :

[0079] in, and These represent the mean-removed cross-section and the Gaussian template, respectively.

[0080] The consistency score of the lateral point spread function Used to measure whether a local bright line structure has the lateral point spread function broadening that a true optical imaging fringe should have.

[0081] The second physical constraint is the longitudinal energy continuity, denoted as lEC. Centered on coordinates (x, y), along the estimated principal direction... Bilinear sampling is used to extract a length of [length missing] from the multi-scale feature map. longitudinal section : ; in, Indicates the sampling radius of the longitudinal profile; According to the longitudinal section Constructing longitudinal energy continuity fractions : .

[0082] Longitudinal energy continuity fraction It is used to measure whether the energy of a target is continuous and stable along the direction of motion. True spatial fringes are formed by the continuous motion integration during exposure and usually have good longitudinal continuity; hot pixel strings, bad columns, or partial diffraction interference may show abrupt intensity changes, local discontinuities, or energy concentrations.

[0083] Consistency score of the lateral point spread function and longitudinal energy continuity fraction Perform weighted fusion to generate a physically consistent attention graph: ; in, They represent learnable scalars, This represents the Sigmoid activation function.

[0084] The physically consistent attention map modulates the multi-scale feature map using a residual weighting method: ; in, Represents the multi-scale feature map after attention modulation; Represents a multi-scale feature map before attention modulation; This indicates element-wise multiplication and is broadcast along the channel dimension.

[0085] Residual connections are used to ensure that even if the physically consistent attention map is unstable in the early stages of training, the module can still maintain a near-identical mapping, thereby improving training stability.

[0086] IV. Lightweight Detection Module for Stripe Direction Sensing The stripe orientation-aware lightweight detection module is used in the network backend to perform stripe orientation-aware feature enhancement and lightweight multi-scale fusion. The module includes a multi-directional reparameterized stripe convolutional block and a high-resolution multi-scale fusion structure. First, the multi-directional reparameterized stripe convolutional block is used to enhance the orientation-aware features of the attention-modulated multi-scale feature map. Then, the high-resolution multi-scale fusion structure is used to fuse the multi-scale features, outputting an enhanced multi-scale feature map. The fusion weights are jointly determined by learnable fusion weights and a physical response gating method after global average pooling of the physical response map.

[0087] The multi-directional reparameterized strip convolutional block adopts a multi-branch structure during the training phase, including a k×k depthwise separable convolutional branch, a 1×k horizontal strip convolutional branch, a k×1 vertical strip convolutional branch, a 1×1 point convolutional branch, and an identity mapping branch. Among them, the k×k depthwise separable convolutional branch is used to extract local features, the 1×k horizontal strip convolutional branch is used to enhance the horizontal stripe response, the k×1 vertical strip convolutional branch is used to enhance the vertical stripe response, the 1×1 point convolutional branch is used for channel fusion, and the identity mapping branch is used to preserve the residual information. Output feature map of multi-directional reparameterized strip convolutional blocks during the training phase for:

[0088] in, This represents the output of the i-th convolutional branch. When i=1, it is the output of the k×k depth separable convolutional branch; when i=2, it is the output of the 1×k horizontal strip convolutional branch; when i=3, it is the output of the k×1 vertical strip convolutional branch; and when i=4, it is the output of the 1×1 point convolutional branch. Indicates the first Batch normalization operation for each convolutional branch This represents the input feature map of the multi-directional reparameterized strip convolutional block at the current scale. During the usage phase, the multi-branch structure from the training phase is folded into a single equivalent convolution through structure reparameterization; for any combination of convolution and batch normalization, the equivalent weights and biases are:

[0089]

[0090] in, Represents the convolution weights. These represent the batch normalization parameters, This represents a constant that prevents division by zero.

[0091] The kernels of the 1×k horizontal strip convolution branch, the k×1 vertical strip convolution branch, and the 1×1 point convolution branch are expanded to k×k size using zero padding. The identity mapping branch is represented as a unit impulse convolution kernel. Finally, the equivalent weights and biases of all branches are summed to obtain a single-path convolution. ; in, This represents the output feature map during the usage phase; This represents the equivalent convolution kernel after fusion; Indicates the convolution operation; This indicates the bias after fusion.

[0092] Therefore, this module has multi-directional expressive capabilities during the training phase and maintains computational complexity similar to that of a single convolution during the usage phase, thus balancing detection accuracy and deployment efficiency.

[0093] The high-resolution multi-scale fusion structure employs a lightweight bidirectional feature pyramid structure, comprising three feature levels: P2, P3, and P4, which are used to receive and process the multi-scale feature maps after physical consistency attention modulation. , .

[0094] For multi-scale features from three levels (P2, P3, and P4), the fusion weights of the lightweight bidirectional feature pyramid structure are jointly determined by the learnable fusion weights and the physical response gating after global average pooling of the physical response map:

[0095]

[0096] in, This represents the enhanced multi-scale fused feature map; This represents the feature map at the i-th scale input to the lightweight bidirectional feature pyramid structure; This represents the fusion weight at the i-th scale; Represents the learnable fusion weights at the i-th scale. This represents the physical response map downsampled to the i-th scale. Indicates a learnable scalar. This indicates global average pooling.

[0097] This design gives higher weight to scale features with strong physical responses during the fusion process, thereby passing the front-end physical response to the back-end multi-scale detection stage.

[0098] Spatial stripe targets are typically extremely narrow and highly sensitive to downsampling. Therefore, the stripe direction-aware lightweight detection module uses only high-resolution scale features stored at three levels (P2, P3, and P4) for fusion, avoiding the use of deeper P5 features to prevent the blurring or loss of weak stripe signals caused by 32x downsampling. Compared to conventional detection networks that use 32x downsampling of deep features, this invention focuses on utilizing the three higher-resolution levels (P2, P3, and P4) in the stripe direction-aware lightweight detection module to reduce information loss in weak, thin stripes during over-downsampling.

[0099] V. Detection Head The detection head can adopt a single-stage target detection structure, which directly predicts the spatial stripe target detection box B for each candidate position, forming a set of spatial stripe target detection boxes.

[0100] Each spatial stripe target detection box B includes: ; in, This indicates the center coordinates of the spatial stripe target detection box. This represents the width and height of the spatial stripe target detection box. This indicates the confidence level of the spatial fringe target.

[0101] For multiple spatial stripe target detection boxes output by the detection head, conventional post-processing methods such as confidence threshold filtering and non-maximum suppression can be used to remove duplicate detection boxes and obtain the final stripe target detection box set.

[0102] In the training phase of the self-supervised stripe detection network, training samples are constructed using a single-label physical injection method. This training method uses unlabeled real background images as a basis, and synthesizes spatial stripe targets within a physically reasonable parameter range by estimating the point spread function width and noise level of the background image, and automatically generates false labels for detection boxes.

[0103] The specific steps are as follows: First, input an unlabeled, realistic background image. In real background image Local intensity maxima are detected as candidate stars, and a two-dimensional Gaussian fit is performed on the neighborhood of the candidate stars to estimate the width of the spread function at the image points. Simultaneously, robust noise standard deviation is calculated in the dark background region to obtain noise estimates. .

[0104] Secondly, in terms of the width of the image point spread function and noise estimates Sampling stripe parameters in the physical distribution centered on:

[0105] in, Indicates the direction of the stripes in this sampling; This indicates the stripe length in this sample; This represents the width of the point spread function in this sampling. This represents the radiant energy accumulated per unit length of the stripe along its trajectory. Indicates the starting point of the stripe.

[0106] Based on the sampled stripe parameters Generate stripe signal .

[0107] Secondly, based on stripe signals Training images were generated using a Poisson electron counting model. :

[0108] in, This represents the Poisson electron counting model; Finally, align the circumscribed rectangle with the calculation axes based on the fringe start and end points, and expand it outwards. Boundary margin, generating unique pseudo-labels :

[0109] in, Indicates the center coordinates of the false label detection box. This indicates the width and height of the pseudo-label detection box.

[0110] This training method only generates pseudo-labels for detection boxes, without generating additional pseudo-labels for orientation, length, mask, centerline, energy continuity, etc. Physical information such as fringe orientation, point spread function width, and energy continuity is internalized by the physical prior encoding module and the physical consistency attention network in the detection network structure, thereby reducing the complexity of pseudo-label construction.

[0111] Compared with the best existing technology, the present invention has the following significant advantages: (1) Reduce reliance on real manually labeled data and reduce data construction costs.

[0112] Existing supervised deep learning methods typically require a large number of real spatial stripe images and manually labeled results. However, real samples of spatial stripe targets are scarce, and the boundary labeling of weak stripes, broken stripes, and blurred stripes is costly and inconsistent.

[0113] This invention uses unlabeled real background images as a basis. By estimating the point spread function width and noise level, it automatically samples physically reasonable stripe parameters and generates synthetic training images. Training can be completed simply by simultaneously generating pseudo-labels for detection boxes. This method significantly reduces the need for manual annotation, lowers the cost of training data construction, and improves the applicability of the method in scenarios where real annotations are insufficient.

[0114] (2) Enhance the detection capability of weak and thin stripes and improve the response stability under low signal-to-noise ratio conditions.

[0115] Traditional Radon transform, Hough transform, and orientation template matching methods rely on fixed templates, manual thresholds, and parameter search. Ordinary convolutional neural networks typically learn stripe features from randomly initialized convolutional kernels. Both are difficult to stably adapt to low signal-to-noise ratio weak stripe scenes.

[0116] This invention transforms the physical imaging principle of observed fringes formed by the convolution of moving line segments and point spread functions into a learnable set of physically parameterized convolutional kernels through a physical prior coding module. This enables the detection network front-end to directly possess an initial response capability consistent with the fringe imaging mechanism. Therefore, this invention can improve the feature response intensity and training efficiency of weak fringes, and reduce the problem of ordinary convolution inefficiently learning the basic fringe morphology from data.

[0117] (3) Improve the ability to suppress bright linear interference and reduce the false alarm rate.

[0118] In actual starry sky images, interference such as cosmic ray hits, hot pixel strings, bad columns, diffraction spikes, and saturated star trails may visually resemble real space streaks. Existing general-purpose depth models and pure data-driven attention mechanisms are prone to misjudging such pseudo-streaks as targets.

[0119] The physical consistency attention module provided by this invention explicitly calculates the lateral point spread function consistency and longitudinal energy continuity based on the estimated principal direction and estimated point spread function width provided by the physical prior coding module. Real motion fringes typically satisfy both lateral optical broadening and longitudinal continuous extension, while many types of pseudo-fringe interference do not possess complete physical consistency. Therefore, this invention can enhance the response of real fringes and suppress pseudo-fringe responses from the perspective of physical generation mechanisms, thereby reducing the false alarm rate in complex backgrounds.

[0120] (4) Better protect high-resolution stripe information and improve detection integrity and positioning reliability.

[0121] General multi-scale detection networks typically use deep, low-resolution features for fusion, but the width of spatial stripe targets is often only a few pixels, and they are easily smoothed, broken, or even lost after over-sampling.

[0122] The lightweight stripe direction sensing detection module provided by this invention employs a high-resolution-biased multi-scale fusion strategy, using only 4x, 8x, and 16x downsampled features. Simultaneously, it combines learnable fusion weights with physical response gating from a physical prior coding module, giving higher weights to scale features with stronger physical responses. This design better preserves the detailed structure of weak stripes, reducing missed detections, breaks, and positioning errors.

[0123] (5) It balances detection accuracy and inference efficiency, making it easy to use in lightweight applications.

[0124] Existing high-precision detection networks typically improve performance by increasing network depth, number of channels, or complex attention structures, which leads to an increase in the number of parameters, computational cost, and inference latency.

[0125] The lightweight detection module for fringe direction sensing provided by this invention employs a reparameterized design with multiple branches in the training phase and a single branch in the inference phase. During training, it enhances the fringe direction representation capability through k×k convolutions, horizontal stripe convolutions, vertical stripe convolutions, point convolutions, and identity mapping branches. In the inference phase, the multiple branches are folded into a single equivalent convolution. Therefore, this invention can reduce the computational overhead in the inference phase while maintaining strong fringe direction sensing capability, thus improving the efficiency of large-scale spatial observation image processing.

[0126] Accordingly, according to embodiments of the present invention, the present invention also provides a computer device and a readable storage medium.

[0127] Figure 2 This is a schematic diagram of the structure of a computer device 12 provided in an embodiment of the present invention. Figure 2 A block diagram of an exemplary computer device 12 suitable for implementing embodiments of the present invention is shown. Figure 2 The computer device 12 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of the present invention.

[0128] like Figure 2 As shown, computer device 12 is represented in the form of a general-purpose computing device. Computer device 12 is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic devices can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0129] The components of the computer device 12 may include, but are not limited to: one or more processors or processing units 16, system memory 28, and bus 18 connecting different system components (including system memory 28 and processing unit 16).

[0130] Bus 18 represents one or more of several bus architectures, including a memory bus or memory controller, a peripheral bus, a graphics acceleration port, a processor, or a local bus using any of the various bus architectures. For example, these architectures include, but are not limited to, the Industry Standard Architecture (ISA) bus, the Micro Channel Architecture (MAC) bus, the Enhanced ISA bus, the Video Electronics Standards Association (VESA) local bus, and the Peripheral Component Interconnect (PCI) bus.

[0131] Computer device 12 typically includes a variety of computer system readable media. These media can be any available media that can be accessed by computer device 12, including volatile and non-volatile media, removable and non-removable media.

[0132] System memory 28 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) 30 and / or cache memory 32. Computer device 12 may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, storage system 34 may be used to read and write non-removable, non-volatile magnetic media (…). Figure 2 Not shown; usually referred to as a "hard drive"). Although Figure 2 Not shown, a disk drive for reading and writing to a removable non-volatile disk (e.g., a "floppy disk") and an optical disk drive for reading and writing to a removable non-volatile optical disk (e.g., a CD-ROM, DVD-ROM, or other optical media) may be provided. In these cases, each drive may be connected to bus 18 via one or more data media interfaces. Memory 28 may include at least one program product having a set (e.g., at least one) of program modules configured to perform the functions of the embodiments of the present invention.

[0133] A program / utility 40 having a set (at least one) of program modules 42 may be stored, for example, in memory 28. Such program modules 42 include, but are not limited to, an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment. Program modules 42 typically perform the functions and / or methods described in the embodiments of the present invention.

[0134] Computer device 12 can also communicate with one or more external devices 14 (e.g., keyboard, pointing device, display 24, etc.), and with one or more devices that enable a user to interact with computer device 12, and / or with any device that enables computer device 12 to communicate with one or more other computing devices (e.g., network card, modem, etc.). This communication can be performed via input / output (I / O) interface 22. Furthermore, computer device 12 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 20. As shown, network adapter 20 communicates with other modules of computer device 12 via bus 18. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with computer device 12, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0135] The processing unit 16 executes various functional applications and data processing by running programs stored in the system memory 28, such as implementing the spatial stripe target detection method provided in the embodiments of the present invention.

[0136] This invention also provides a non-transitory computer-readable storage medium storing computer instructions, on which a computer program is stored, wherein when the program is executed by a processor, the processor performs the spatial stripe target detection method provided in any embodiment of this application.

[0137] The computer storage medium of this invention can be any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. More specific examples (a non-exhaustive list) of computer-readable storage media include: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0138] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0139] The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof. The computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer can be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0140] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0141] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A spatial stripe target detection method based on a physics prior-driven self-supervised network, characterized in that, include: S1: Establish a self-supervised fringe detection network based on physical priors, which includes a physical prior encoding module, a multi-scale feature extraction module, a physical consistency attention module, a fringe direction perception lightweight detection module, and a detection head; S2: Acquire spatial observation images and input them into the physical prior coding module. Based on the physical parameterized convolutional kernel group, the physical imaging prior of the spatial stripe target is explicitly encoded to generate a physical response map. S3: Input the physical response map into the multi-scale feature extraction module to extract at least three multi-scale feature maps of different scales; S4: Input the multi-scale feature map into the physical consistency attention module, determine the estimated principal direction and the estimated point spread function width based on the physical response map and the physical parameter set corresponding to the response in each direction, calculate the horizontal point spread function consistency score and the vertical energy continuity score respectively, fuse them to generate the physical consistency attention map, and modulate the multi-scale feature map; S5: Input the attention-modulated multi-scale feature map into the stripe direction-aware lightweight detection module, enhance the direction-aware features through multi-directional reparameterized stripe convolutional blocks, and perform multi-scale feature fusion using a high-resolution multi-scale fusion structure to output the enhanced multi-scale feature map; the fusion weights are jointly determined by the learnable fusion weights and the physical response gating based on the physical response map. S6: Input the enhanced multi-scale feature map into the detection head to predict the spatial stripe target detection box.

2. The spatial stripe target detection method based on a physics prior-driven self-supervised network according to claim 1, characterized in that, The physical prior coding module defines a set of physical parameters consisting of M directional anchor points: ； in, This represents the direction angle of the m-th anchor point. This represents the length of the m-th stripe template. This represents the width of the point spread function of the m-th stripe template; Generate a set of physically parameterized convolutional kernels based on the set of physical parameters.

3. The spatial stripe target detection method based on a physics prior-driven self-supervised network according to claim 2, characterized in that, The m-th stripe template in the physically parameterized convolution kernel group is generated as follows: ； ； in, This represents the fringe template response value at coordinates (x, y) in the spatial observation image; Represents the normalization constant; Represents the coordinates along the principal direction of the stripes; Represents the coordinates perpendicular to the principal direction of the stripes; Using the Sigmoid function for rectangle functions Smoothing is performed to make the stripe template aligned with the direction angle. Stripe template length and point spread function width Preserving differentiability: in, This represents the Sigmoid activation function. This represents the hyperparameter used to control the width of the transition band.

4. The spatial stripe target detection method based on a physics prior-driven self-supervised network according to claim 3, characterized in that, The physical response diagram is generated as follows: For an input spatial observation image I, the orientation response map generated by the m-th fringe template for: ； in, Represents a two-dimensional convolution operation; The M directional response maps are concatenated along the channel dimension and then mixed using a 1×1 convolution to obtain the physical response map: ； in, This represents the 1×1 convolution weight.

5. The spatial stripe target detection method based on a physics prior-driven self-supervised network according to claim 1, characterized in that, The multi-scale feature extraction module extracts a first-scale feature map with a downsampling factor of 4, a second-scale feature map with a downsampling factor of 8, and a third-scale feature map with a downsampling factor of 16 from the physical response map. The first-scale feature map, the second-scale feature map, and the third-scale feature map together form a multi-scale feature map.

6. The spatial stripe target detection method based on a physics prior-driven self-supervised network according to claim 1, characterized in that, The physical consistency attention module determines the estimated principal direction based on the physical response map and the physical parameter set corresponding to the response in each direction. And the estimated point spread function width is ; The consistency score of the lateral point spread function is calculated as follows: Centered on coordinates (x, y), along a direction perpendicular to the estimated principal direction In the direction, a length of [value] is extracted from the multi-scale feature map through bilinear sampling. cross section : ； in, Indicates along the direction perpendicular to the estimated principal direction The offset; Indicates the sampling radius of the transverse profile; Represents the feature map at the s-th scale; Construction and width of point spread function Corresponding Gaussian template : The lateral profile is normalized and matched with the Gaussian template to obtain the consistency score of the lateral point spread function. : in, and These represent the mean-removed cross-section and the Gaussian template, respectively. The longitudinal energy continuity fraction is calculated as follows: Centered on coordinates (x, y), along the estimated principal direction Bilinear sampling is used to extract a length of [length missing] from the multi-scale feature map. longitudinal section : ； in, Indicates the sampling radius of the longitudinal profile; According to the longitudinal section Constructing longitudinal energy continuity fractions : 。 7. The spatial stripe target detection method based on a physics prior-driven self-supervised network according to claim 6, characterized in that, The lateral point spread function consistency score and the longitudinal energy continuity score are weighted and fused to generate a physically consistent attention map: ； in, They represent learnable scalars, This represents the Sigmoid activation function; The physically consistent attention map modulates the multi-scale feature map using a residual weighting method: ； in, This represents the modulated multi-scale feature map; Represents the multi-scale feature map before modulation; This indicates element-wise multiplication.

8. The spatial stripe target detection method based on a physics prior-driven self-supervised network according to claim 1, characterized in that, The multi-directional reparameterized strip convolutional block adopts a multi-branch structure during the training phase, including a k×k depthwise separable convolutional branch, a 1×k horizontal strip convolutional branch, a k×1 vertical strip convolutional branch, a 1×1 point convolutional branch, and an identity mapping branch. Among them, the k×k depthwise separable convolutional branch is used to extract local features, the 1×k horizontal strip convolutional branch is used to enhance the horizontal stripe response, the k×1 vertical strip convolutional branch is used to enhance the vertical stripe response, the 1×1 point convolutional branch is used for channel fusion, and the identity mapping branch is used to preserve the residual information. Output feature map of multi-directional reparameterized strip convolutional blocks during the training phase for: in, This represents the output of the i-th convolutional branch. When i=1, it is the output of the k×k depth separable convolutional branch; when i=2, it is the output of the 1×k horizontal strip convolutional branch; when i=3, it is the output of the k×1 vertical strip convolutional branch; and when i=4, it is the output of the 1×1 point convolutional branch. Indicates the first Batch normalization operation for each convolutional branch This represents the input feature map of the multi-directional reparameterized strip convolutional block at the current scale. During the usage phase, the multi-branch structure from the training phase is folded into a single equivalent convolution through structure reparameterization; for any combination of convolution and batch normalization, the equivalent weights and biases are: in, Represents the convolution weights. These represent the batch normalization parameters, A constant that prevents division by zero; The kernels of the 1×k horizontal strip convolution branch, the k×1 vertical strip convolution branch, and the 1×1 point convolution branch are expanded to k×k size using zero padding. The identity mapping branch is represented as a unit impulse convolution kernel. Finally, the equivalent weights and biases of all branches are summed to obtain a single-path convolution. ； in, This represents the output feature map during the usage phase; This represents the equivalent convolution kernel after fusion; Indicates the convolution operation; This indicates the bias after fusion.

9. The spatial stripe target detection method based on a physics prior-driven self-supervised network according to claim 1, characterized in that, The high-resolution multi-scale fusion structure adopts a lightweight bidirectional feature pyramid structure. The fusion weights of the lightweight bidirectional feature pyramid structure are determined as follows: in, This represents the enhanced multi-scale fused feature map; This represents the feature map at the i-th scale input to the lightweight bidirectional feature pyramid structure; This represents the fusion weight at the i-th scale; Represents the learnable fusion weights at the i-th scale. This represents the physical response map downsampled to the i-th scale. Indicates a learnable scalar. This indicates global average pooling.

10. The spatial stripe target detection method based on a physics prior-driven self-supervised network according to claim 1, characterized in that, During the training phase of the self-supervised stripe detection network, training samples are constructed using a single-label physical injection method, specifically including: First, input an unlabeled, realistic background image. In real background image Local intensity maxima are detected as candidate stars, and a two-dimensional Gaussian fit is performed on the neighborhood of the candidate stars to estimate the width of the spread function at the image points. Simultaneously, robust noise standard deviation is calculated in the dark background region to obtain noise estimates. . Secondly, in terms of the width of the image point spread function and noise estimates Sampling stripe parameters in the physical distribution centered on: in, Indicates the direction of the stripes in this sampling; This indicates the stripe length in this sample; This represents the width of the point spread function in this sampling. This represents the radiant energy accumulated per unit length of the stripe along its trajectory. Indicates the starting point of the stripes; Based on the sampled stripe parameters Generate stripe signal ; Secondly, based on stripe signals Training images were generated using a Poisson electron counting model. : in, This represents the Poisson electron counting model; Finally, align the circumscribed rectangle with the calculation axes based on the fringe start and end points, and expand it outwards. Boundary margin, generating unique pseudo-labels : in, Indicates the center coordinates of the false label detection box. This indicates the width and height of the pseudo-label detection box.

11. The spatial stripe target detection method based on a physics prior-driven self-supervised network according to claim 1, characterized in that, Set of spatial stripe target detection boxes: ； in, This indicates the center coordinates of the spatial stripe target detection box. This represents the width and height of the spatial stripe target detection box. This indicates the confidence level of the spatial fringe target.

12. A computer device, characterized in that, include: At least one processor; as well as A memory that is communicatively connected to at least one processor; wherein, The memory stores instructions that can be executed by at least one processor, which, when executed by at least one processor, causes the at least one processor to perform the spatial stripe target detection method according to any one of claims 1 to 11.

13. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to execute the spatial stripe target detection method according to any one of claims 1 to 11.

Images