A Daytime Star Detection Method Based on Spatiotemporal Feature Context Enhancement

Abstract

This invention discloses a daytime star detection method based on spatiotemporal feature context enhancement, comprising: designing a frame difference-optical flow guided spatiotemporal feature context enhancement module; extracting spatial features, frame difference features, and optical flow motion features of star targets in consecutive frames through a sliding observation window; supplementing these features to a data-driven network through multi-flow feature fusion; constructing a synthetic star map dataset covering various noise and background distributions; using U-Net as the backbone network to implement an encoding and decoding architecture for star segmentation training; and verifying the performance of the method through simulation experiments and field star observation experiments. This invention expands the receptive field of stars in the temporal dimension, enhances the features of faint stars, and suppresses non-stationary background interference, significantly improving the detection accuracy and robustness of faint stars against complex daytime backgrounds, and is suitable for daytime faint star detection on all-day star sensors.

Description

Technical Field

[0001] This invention relates to the field of all-day star sensor technology, specifically to a method for detecting faint stars under daytime background noise interference. Background Technology

[0002] Star sensors are a crucial technology and core product for autonomous attitude measurement of spacecraft, and have been widely used in various satellites, spacecraft, deep space probes, and other aerospace fields. All-weather star sensors further solve the problem of traditional star sensors not being able to operate at all times in near-Earth space, effectively expanding the application range of astronomical navigation. However, when observing stars during the day, strong atmospheric background interference, optical device defects, and detector noise can all reduce the signal-to-noise ratio of stellar targets, triggering numerous false alarms and severely affecting the stability and accuracy of attitude determination.

[0003] Traditional methods for detecting faint stars primarily rely on spatial features (such as the Gaussian distribution and local brightness characteristics of stars). However, in environments with strong noise interference, the spatial feature differences between stars and quasar noise become blurred, making effective differentiation difficult. Some methods attempt to improve detection performance by introducing temporal dimension features. Among existing related technical solutions, patent document CN117635530A discloses a star detection method based on multi-scale feature fusion, which enhances the spatial feature representation of stars by constructing a multi-scale feature extraction network. However, this method does not fully utilize temporal information and is susceptible to transient noise interference in complex dynamic backgrounds, resulting in insufficient detection rate for faint stars. Patent document CN119478447A proposes a star map denoising method combining temporal difference. Although it introduces temporal dimension information, it mainly relies on the difference operation between adjacent frames to suppress static backgrounds. For faint stars with slight displacements between frames, the difference operation can easily lead to target signal attenuation.

[0004] In related technical fields, Jinming Du et al. proposed a spatial-temporal feature-based detection framework for infrared dim small target (IEEE Transactions on Geoscience and RemoteSensing, 2021) in 2021. This framework achieves target detection by fusing local spatial contrast and temporal motion information. However, its spatial-temporal feature fusion strategy is not optimized for the motion characteristics of star points, making it prone to false detections in scenes with complex stellar backgrounds. Xuyang Liu et al. proposed the IR-MPE method (IR-MPE: A Long-Term Optical Flow Based Motion Pattern Extractor for Infrared Small Dim Targets, IEEE Transactions on Instrumentation and Measurement, 2024) in 2024. This method extracts the motion patterns of infrared small targets based on long-term optical flow. However, this method requires high continuity of the target's trajectory. In star point observation, atmospheric turbulence often causes abrupt changes in the trajectory, leading to decreased stability of feature extraction and making it difficult to adapt to the dynamic scene requirements of star point detection.

[0005] Therefore, there is an urgent need for a detection method that can integrate spatiotemporal features and enhance the representation of faint star points, and can actively guide the network to focus on star point targets with relative displacement between frames. Summary of the Invention

[0006] To address the issues of low accuracy and poor anti-interference capability in detecting faint star points under strong background noise, this invention provides a faint star point detection method based on spatiotemporal feature context enhancement. By fusing spatial features and temporal motion features, the robustness of faint star point recognition is improved.

[0007] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0008] A daytime star detection method based on spatiotemporal feature context enhancement, characterized in that,

[0009] Step S1: Design the spatiotemporal feature context enhancement module:

[0010] S1.1. Construct a sliding observation window to cache continuous frame sequences. The window expands symmetrically with the current frame as the center and is dynamically updated to preserve temporal coherence.

[0011] S1.2. Feature extraction via three parallel pathways, including the following:

[0012] Original spatial path: The current frame is processed by a convolutional encoder to output spatial structure features;

[0013] Frame difference feature path: calculate the bidirectional inter-frame difference within the window, suppress the background through a truncation function, and then output time-varying features through a convolutional encoder;

[0014] Optical flow motion path: Estimate sparse optical flow and enhance trajectory continuity through gradient constant constraints, outputting motion characteristics;

[0015] S1.3. Segment the features of the three channels along the channel dimension to achieve multi-flow feature fusion;

[0016] Step S2: Construct the dataset and train the network, including:

[0017] S2.1 Obtain star position and brightness level information based on real star catalogs, and generate noise-free star point images by combining star sensor parameters with star imaging and star detection models;

[0018] S2.2. Superimpose real atmospheric background radiation, detector noise, and stray light interference to generate a noisy star map;

[0019] S2.3. Employ data augmentation strategies to balance positive and negative samples, generating training and validation sets with binary segmentation labels;

[0020] S2.4. The U-Net architecture is adopted, and the first three feature encoding processes are retained to reduce the loss of features of small targets;

[0021] S2.5. Embed channel attention mechanism and spatial attention mechanism in the feature fusion stage to enhance the discrimination power of weak star point features;

[0022] S2.6. The network is trained on a synthetic dataset using Focal Loss as the loss function and Adam as the optimizer, with the learning rate dynamically adjusted.

[0023] Step S3: Perform faint star point segmentation based on the trained network, verify the segmentation performance through simulated star map experiments and real star map experiments in the field, and evaluate the effectiveness of the method by trajectory integrity and detection accuracy.

[0024] This invention also discloses a daytime star detection device based on spatiotemporal feature context enhancement, characterized by:

[0025] The spatiotemporal feature context enhancement module is used to construct a sliding observation window to cache continuous frame sequences. The window expands symmetrically with the current frame as the center and is dynamically updated to preserve temporal coherence.

[0026] The feature extraction module, which uses a three-path parallel approach, includes the following components:

[0027] Original spatial path: The current frame is processed by a convolutional encoder to output spatial structure features;

[0028] Frame difference feature path: calculate the bidirectional inter-frame difference within the window, suppress the background through a truncation function, and then output time-varying features through a convolutional encoder;

[0029] Optical flow motion path: Estimate sparse optical flow and enhance trajectory continuity through gradient constant constraints, outputting motion characteristics;

[0030] Features from three channels are spliced ​​along the channel dimension to achieve multi-flow feature fusion;

[0031] The construction of the dataset and training network modules includes:

[0032] Based on real star catalogs, star position and brightness level information are obtained, and noise-free star point images are generated by combining star sensor parameters with star imaging and star detection models.

[0033] By superimposing real atmospheric background radiation, detector noise, and stray light interference, a noisy star map is generated.

[0034] A data augmentation strategy is used to balance positive and negative samples, generating training and validation sets with binary segmentation labels;

[0035] The U-Net architecture is adopted, and the first three feature encoding processes are retained to reduce the loss of features for small targets;

[0036] In the feature fusion stage, channel attention mechanism and spatial attention mechanism are embedded to enhance the discrimination power of weak star point features;

[0037] The network is trained on a synthetic dataset using Focal Loss as the loss function and Adam as the optimizer, with the learning rate dynamically adjusted.

[0038] The faint star point segmentation module verifies its segmentation performance through simulated star map experiments and real-world star map experiments, and evaluates the effectiveness of the method by assessing trajectory integrity and detection accuracy.

[0039] The present invention also relates to an electronic device, the electronic device comprising:

[0040] At least one processor; and,

[0041] A memory communicatively connected to the at least one processor; wherein,

[0042] The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method.

[0043] The present invention also relates to a non-transitory computer-readable storage medium that stores computer instructions for causing the computer to perform the method described thereon.

[0044] Beneficial effects:

[0045] This invention effectively detects low signal-to-noise ratio (SNR) star targets during the day. By designing a temporal feature enhancement module, this module actively learns the motion correlation of star points between consecutive frames. This design not only effectively expands the temporal receptive field of star points but also significantly enhances the feature differences between faint star points and background noise by capturing inter-frame dynamic information, thus solving the problem of blurred features and indistinguishability in traditional methods. Furthermore, this temporal feature enhancement module does not use traditional 3D convolutional layers with large parameter scales. Instead, it introduces a small number of learnable parameters to achieve spatiotemporal feature correlation. This lightweight design improves detection performance while adding very few additional parameters, perfectly adapting to computationally limited applications such as star sensors and avoiding performance bottlenecks caused by insufficient computing power. In addition, because this invention innovatively combines the spatial distribution characteristics and temporal motion characteristics of star points for dual feature constraints, it can accurately capture and identify faint star points even in low SNR environments and complex background interference, thus maintaining a stable high precision and recall rate and effectively reducing the impact of false alarm rate on attitude determination stability. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating the faint star detection method based on spatiotemporal feature context enhancement provided in an embodiment of the present invention.

[0047] Figure 2 This is a schematic diagram of the spatiotemporal motion characteristics of a star point target in the spatiotemporal feature-enhanced faint star point detection method provided in an embodiment of the present invention.

[0048] Figure 3 This is a schematic diagram of the spatiotemporal feature context enhancement module in the faint star point detection method based on spatiotemporal feature context enhancement provided in an embodiment of the present invention.

[0049] Figure 4 A schematic diagram illustrating the process of creating a dataset according to an embodiment of the present invention.

[0050] Figure 5 This is a diagram of a spatiotemporal feature-based context enhancement network structure provided according to an embodiment of the present invention.

[0051] Figure 6 This is a diagram showing the actual star map detection results obtained according to an embodiment of the present invention.

[0052] Figure 7 This is a diagram showing the results of star trajectory detection in a star map sequence obtained according to an embodiment of the present invention. Detailed Implementation

[0053] To enhance the spatiotemporal feature representation of star targets, this invention constructs a sliding observation window to extract the spatial structure, time-varying frame difference, and optical flow motion features of consecutive frames. These features are then supplemented into a symmetric encoder-decoder segmentation network with an attention mechanism through multi-stream fusion, thereby achieving spatiotemporal enhancement and noise suppression of faint star features. Ultimately, this improves the detection performance of faint star targets under daytime background noise interference.

[0054] This invention discloses a daytime star detection method based on spatiotemporal feature context enhancement, see [link to relevant documentation]. Figure 1-2 As shown, it includes the following steps:

[0055] Step S1: Design the spatiotemporal feature context enhancement module:

[0056] The frame star map data is enhanced with temporal features. This invention actively guides the network to enhance the spatiotemporal contextual feature representation of faint stars based on the motion features of stars themselves in multi-frame star maps. The specific steps are as follows.

[0057] 1.1 Construct a sliding observation window to cache continuous frame sequences. The window expands symmetrically with the current frame as the center and is dynamically updated to preserve temporal coherence.

[0058] like Figure 3 As shown, window length (Centered on the current frame, expand forward and backward) (frame), specifically represented as During the update, redundant observation frames outside the window are removed and new observation-related frames are added, preserving temporal coherence.

[0059] 1.2 Three-path feature extraction of the original spatial path: The original spatial path is extracted from the current frame through 3 layers of convolution (kernel size 3×3). Spatial features Frame difference feature path: Calculate the frame difference map of the star map sequence as follows: ,in The frame difference step size, To suppress non-stationary background interference, the inter-frame difference map is convolved to obtain features. Optical flow motion path: The motion vector is obtained using a sparse optical flow estimation algorithm. Calculate the amplitude The motion optical flow map is convolved to obtain features. .

[0060] 1.3 The features of the three channels are spliced ​​along the channel dimension to achieve multi-flow feature fusion.

[0061] splicing along the channel dimension ( , , ), to obtain fusion features This serves as the input for subsequent networks.

[0062] Step S2: Construct the dataset and train the network, specifically including:

[0063] 2.1. For example Figure 4 As shown, stellar position and magnitude information data are obtained based on real star catalogs, and a simulated star map is generated using a stellar imaging model and a detection model. The stellar imaging model is based on the angular rate of a star sensor. , , and focal length The position of the star point in the image coordinate system is:

[0064]

[0065] in , , These are the initial position parameters.

[0066] Star point energy model: Total energy of stars received by the detector:

[0067] in for The light flux of stars of equal magnitude For the transmittance of the optical system, For optical lens diameter, For the quantum efficiency of the detector The exposure time is used, and the pixel grayscale value is obtained by energy quantization, which conforms to a two-dimensional Gaussian distribution.

[0068] 2.2 Superimposing real atmospheric background radiation, detector noise and stray light interference: Poisson-Gaussian noise caused by shot noise, dark current noise and readout noise is superimposed; daytime stray light images containing optical device and detector defects are captured by real cameras and synthesized with simulated star points to improve the realism of the dataset.

[0069] 2.3 Data Augmentation and Label Generation Augmentation Strategies: Randomly rotate star points and copy and paste them to the background area to improve the ratio of positive and negative samples; the label format is a binary segmentation image, with the target pixel area of ​​the star point marked as 1 (pixels covering 90% of the star point energy) and the background pixels marked as 0.

[0070] 2.4 U-Net network architecture is adopted: such as Figure 5 As shown, the first three feature encoding processes are retained to reduce the loss of features for small targets;

[0071] 2.5 Channel attention and spatial attention mechanisms are embedded in the feature fusion stage to enhance the discrimination power of weak star point features; Focal Loss is used as the loss function, and the network is trained on a synthetic dataset. The optimizer is Adam, and the learning rate is dynamically adjusted. The initial learning rate is 0.0001, and it is halved every 20 rounds. Training environment: NVIDIA RTX 4070 GPU, 100 iterations, batch size 10.

[0072] Step S3: Perform faint star point segmentation based on the trained network, verify the segmentation performance through simulated star map experiments and real star map experiments in the field, and evaluate the effectiveness of the method by trajectory integrity and detection accuracy.

[0073] Simulation Experiment

[0074] To evaluate the performance of the method of this invention, experiments were conducted using both simulated star charts and real images. The parameters of the simulated and real star sensors are shown in Table 1. The star sensor has a field of view of 0.4° × 0.4°, an imaging device resolution of 512 × 640 pixels, an observation band of 1300-1700 nm, and a pixel size of 20. m×20 m, the optical system focal length is approximately 1470mm.

[0075] Table 1. All-day star sensor parameters

[0076]

[0077] Eight experiments with different combinations of atmospheric radiation intensity and exposure time were designed to compare model-driven methods (Multi-Scale Local Contrast Method CMLCM and Background Estimation Method BEPF) and data-driven methods (Star-Unet and 3D-Unet). The results show that the present invention achieves a precision of 82.13% and a recall of 89.61% at an SNR of 1.5–2, significantly outperforming the comparative methods.

[0078] 3.2 Field Experiment

[0079] like Figure 6 The image shown is a real daytime star map taken at 7:20 am in Lijiang, Yunnan (altitude 2400m), along with the network segmentation results. Figure 7 The diagram shows the trajectory continuity representation for evaluating detection performance. It can be seen that the star point trajectory of the present invention has high integrity and low false alarm rate, verifying its effectiveness in real-world scenarios.

[0080] The above results demonstrate that the method of the present invention can achieve high accuracy and recall for detecting faint star points under strong background noise interference during the day.

Claims

1. A daytime star detection method based on spatiotemporal feature context enhancement, characterized in that, Step S1: Design the spatiotemporal feature context enhancement module: S1.

1. Construct a sliding observation window to cache continuous frame sequences. The window expands symmetrically with the current frame as the center and is dynamically updated to preserve temporal coherence. S1.

2. Feature extraction via three parallel pathways, including the following: Original spatial path: The current frame is processed by a convolutional encoder to output spatial structure features; Frame difference feature path: calculate the bidirectional inter-frame difference within the window, suppress the background through a truncation function, and then output time-varying features through a convolutional encoder; Optical flow motion path: Estimate sparse optical flow and enhance trajectory continuity through gradient constant constraints, outputting motion characteristics; S1.

3. Segment the features of the three channels along the channel dimension to achieve multi-flow feature fusion; Step S2: Construct the dataset and train the network, including: S2.1 Obtain star position and brightness level information based on real star catalogs, and generate noise-free star point images by combining star sensor parameters with star imaging and star detection models; S2.

2. Superimpose real atmospheric background radiation, detector noise, and stray light interference to generate a noisy star map; S2.

3. Employ data augmentation strategies to balance positive and negative samples, generating training and validation sets with binary segmentation labels; S2.

4. The U-Net architecture is adopted, and the first three feature encoding processes are retained to reduce the loss of features of small targets; S2.

5. Embed channel attention mechanism and spatial attention mechanism in the feature fusion stage to enhance the discrimination power of weak star point features; S2.

6. The network is trained on a synthetic dataset using Focal Loss as the loss function and Adam as the optimizer, with the learning rate dynamically adjusted. Step S3: Perform faint star point segmentation based on the trained network, verify the segmentation performance through simulated star map experiments and real star map experiments in the field, and evaluate the effectiveness of the method by trajectory integrity and detection accuracy.

2. The method according to claim 1, characterized in that, In step S1, the length of the sliding observation window is The current frame is At that time, the window is represented as When the window is updated, old frames outside the observation window are removed and new frames are added; among them, Indicates the observation window of the star chart sequence. Indicates the current observation frame. Star chart sequence observation window sequence length, Indicates the first in the sequence frame.

3. The method according to claim 1, characterized in that, In step S1, the bidirectional inter-frame difference of the frame difference feature path is calculated as follows: ,in For inter-frame difference images, The frame difference step size, Indicates the current observation frame. Indicates the current observation frame after frame, This is a truncation function to suppress non-stationary background interference.

4. The method according to claim 1, characterized in that, In step S1, the optical flow amplitude of the optical flow path is calculated as follows: Where A is the optical flow amplitude diagram, For optical flow vectors and motion vectors, and These are the horizontal and vertical vectors, respectively.

5. The method according to claim 1, characterized in that, In step S2, the channel attention mechanism aggregates channel information through global average pooling and max pooling, and outputs channel weights through MLP; the spatial attention mechanism outputs spatial weights through channel-dimensional pooling and convolution.

6. A daytime star detection device based on spatiotemporal feature context enhancement, characterized by: The spatiotemporal feature context enhancement module is used to construct a sliding observation window to cache continuous frame sequences. The window expands symmetrically with the current frame as the center and is dynamically updated to preserve temporal coherence. The feature extraction module, which uses a three-path parallel approach, includes the following components: Original spatial path: The current frame is processed by a convolutional encoder to output spatial structure features; Frame difference feature path: calculate the bidirectional inter-frame difference within the window, suppress the background through a truncation function, and then output time-varying features through a convolutional encoder; Optical flow motion path: Estimate sparse optical flow and enhance trajectory continuity through gradient constant constraints, outputting motion characteristics; Features from three channels are spliced ​​along the channel dimension to achieve multi-flow feature fusion; The construction of the dataset and training network modules includes: Based on real star catalogs, star position and brightness level information are obtained, and noise-free star point images are generated by combining star sensor parameters with star imaging and star detection models. By superimposing real atmospheric background radiation, detector noise, and stray light interference, a noisy star map is generated. A data augmentation strategy is used to balance positive and negative samples, generating training and validation sets with binary segmentation labels; The U-Net architecture is adopted, and the first three feature encoding processes are retained to reduce the loss of features for small targets; In the feature fusion stage, channel attention mechanism and spatial attention mechanism are embedded to enhance the discrimination power of weak star point features; The network is trained on a synthetic dataset using Focal Loss as the loss function and Adam as the optimizer, with the learning rate dynamically adjusted. The faint star point segmentation module verifies its segmentation performance through simulated star map experiments and real-world star map experiments, and evaluates the effectiveness of the method by assessing trajectory integrity and detection accuracy.

7. An electronic device, characterized in that, The electronic device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-5.

8. A non-transitory computer-readable storage medium storing computer instructions for causing the computer to perform the method of any one of claims 1-5.

Images