Dark and weak space small target detection method based on airspace guidance

Abstract

The invention belongs to the field of space target detection, and relates to a weak space small target detection method based on airspace guidance, which comprises the following steps: inputting a to-be-detected space target image sequence into a pre-established and trained weak space small target time sequence evolution feature sensing network based on airspace guidance to obtain a target prediction frame, small target detection in a dark space is realized; the dark space small target time sequence evolution feature sensing network extracts initial time-space features of an image sequence through a lightweight three-dimensional convolutional network, enhances target response and inhibits star clutters, models target time sequence evolution features and enhances target space region features by constructing a double-branch structure, and finally, the target time sequence evolution feature sensing network is constructed. A gating-based double attention feature fusion module is adopted to selectively enhance and fuse features output by the branch structure, and the joint expression ability of position and motion information is improved.

Description

Technical Field

[0001] The present invention belongs to the field of space target detection, and in particular relates to a method for detecting small targets in dim space based on spatial guidance. Background Art

[0002] Faint and small space targets primarily refer to space targets (such as satellites, spacecraft, and space debris) detected in long-range space-based passive optical observations. Due to their distance from the observation platform, small size, and weak signal strength, they occupy only a very small pixel area on the imaging plane and have a low signal-to-noise ratio. These targets appear dim, weak, and small in the image, making them difficult to discern against a complex background. In recent years, with the rapid development of aerospace technology, space situational awareness (SSA) has gradually become a research focus worldwide. Faint and small space target detection, as a key technology in SSA systems, holds significant application value in satellite surveillance, space security warning, orbit maintenance, and space defense. However, detecting faint and small space targets in complex deep space environments with low SNR conditions remains a significant challenge.

[0003] Existing methods for faint target detection can be broadly categorized as model-driven and data-driven. For the past few decades, faint small target detection has primarily relied on model-driven methods, employing manually designed detectors to identify targets based on significant feature differences between the target and the background. Typical methods include ephemeris calculation, Hough transforms, multi-frame temporal projections, and tracking before detection. In addition, there are methods based on filters, the human visual system, and low-rank sparse decomposition. These methods are typically based on mature mathematical theories and offer good interpretability in target feature modeling. However, faced with strong stellar clutter and strong noise interference in complex deep space environments, low-signal-to-noise ratio targets are easily overwhelmed by the background, making it difficult for traditional methods to effectively identify targets and significantly reducing detection performance.

[0004] In recent years, the rapid development of convolutional neural networks has led to significant progress in the field of object detection, providing new solutions for detecting faint targets in low signal-to-noise ratio (SNR) conditions. Leveraging the powerful feature representation and learning capabilities of deep learning, data-driven methods can automatically extract target features from large datasets, effectively improving detection accuracy and gradually becoming a key research direction in faint, small target detection. Current data-driven methods can be categorized as single-frame and multi-frame image-based. Single-frame detection methods primarily rely on the target's grayscale distribution, learning local spatial features through neural networks to locate the target. This approach offers the advantages of low computational complexity and fast detection speed. However, in complex deep-space backgrounds, targets are often faint and exhibit similar spatial characteristics to background stellar clutter, leading to serious false alarm and missed detection issues with single-frame detection. In contrast, multi-frame detection methods further incorporate target motion information, leveraging the target's distribution characteristics in both temporal and spatial dimensions to extract more discriminative features and improve the detectability of faint, small targets. However, existing multi-frame detection methods are primarily designed for scenes with stable and textured backgrounds, such as the ground or sky, for example, infrared drone target detection. These methods usually process the target's spatial and temporal features separately, ignoring the synergistic relationship between the two, and are difficult to adapt to deep space exploration scenarios with target-like celestial bodies with high clutter, weak texture features and strong background interference. Summary of the Invention

[0005] The purpose of the present invention is to overcome the defects of the prior art and propose a small target detection method in dark space based on spatial guidance.

[0006] In view of this, the present invention proposes a method for detecting small targets in dim space based on spatial guidance, comprising: The image sequence of the space target to be detected is input into the pre-established and trained spatial domain-guided dim space small target temporal evolution feature perception network to obtain the target prediction frame and realize dim space small target detection; The proposed network for perceiving the temporal evolution of small targets in dim space extracts the initial spatiotemporal features of image sequences through a lightweight three-dimensional convolutional network, enhances target response and suppresses stellar clutter. By constructing a dual-branch structure, it models the temporal evolution features of targets and enhances the spatial regional features of targets. A gated dual-attention feature fusion module is used to selectively enhance and fuse the features output by the branch structure, thereby improving the joint expression capability of position and motion information.

[0007] Preferably, the input of the dim space small target temporal evolution feature perception network is a continuous 5-frame image sequence, and the output is the target prediction frame of the last frame image of the current image sequence. The dim space small target temporal evolution feature perception network includes: an initial feature extraction module, a spatiotemporal dual-branch feature extraction module, a gated dual attention feature fusion module and a target detection head, wherein, The initial feature extraction module is used to extract 5 consecutive frames of images of the space target to be detected. Perform initial spatiotemporal feature extraction to obtain the corresponding initial spatiotemporal features ; The spatiotemporal dual-branch feature extraction module is used to extract the target time domain and spatial domain features respectively through the temporal evolution feature perception branch and the spatial region enhancement branch based on cross attention, and obtain the time domain features. and spatial characteristics ; The gate-based dual attention feature fusion module is used to fusion the temporal features and spatial characteristics Fusion is performed to obtain the target spatiotemporal correlation features ; The target detection head is used to detect the temporal and spatial correlation features of the target Processing is performed to obtain the final target prediction box result.

[0008] Preferably, the initial feature extraction module includes: a sequentially connected time dimension stack, three groups of alternating series-connected three-dimensional convolution blocks and downsampling, and two two-dimensional convolution blocks; wherein the three-dimensional convolution block includes a sequentially connected three-dimensional convolution layer, batch normalization, and a nonlinear activation function; the downsampling includes a sequentially connected three-dimensional convolution layer with a stride of 2, batch normalization, and a nonlinear activation function, and the two-dimensional convolution block includes a sequentially connected two-dimensional convolution layer, batch normalization, and a nonlinear activation function; The specific processing process includes: Input a sequence of 5 consecutive image frames ,in, is the reference frame, is a key frame, and the space-time tensor is obtained by stacking in the time dimension ,in, is the number of channels of the input image, is the number of image rows, is the number of image columns; After three sets of alternating series of three-dimensional convolution blocks and downsampling, the target spatiotemporal features are obtained. , perform feature splitting in the time dimension and obtain the i Spatiotemporal features corresponding to frame images ,in ; Will The loop inputs two 2D convolution blocks in series and obtains the i Initial spatiotemporal characteristics of frame images .

[0009] Preferably, the spatial region enhancement branch based on cross attention includes: feature connection, two serially connected two-dimensional convolution blocks, three parallel two-dimensional convolution blocks, cross attention, two-dimensional convolution block, three parallel two-dimensional convolution blocks and parameter normalization. The specific processing process includes: Initial spatiotemporal characteristics of the reference frame First, the channel dimension feature connection is performed, and the spatial difference features of the reference frame are further extracted through two serial two-dimensional convolution blocks. The query Q corresponding to the key frame, the key K and value V corresponding to the reference frame are generated through three parallel two-dimensional convolution blocks respectively; based on the attention mechanism, the key frame spatial features weighted by the cross attention coefficient are obtained ; The initial spatiotemporal features of the keyframes and Perform residual connection, and extract the multi-scale features of the target area through three parallel two-dimensional convolution blocks. Then, add the multi-scale features and obtain the spatial weighting coefficient after parameter normalization. Then, weight the key frame features and output the spatial domain features. .

[0010] Preferably, the temporal evolution feature perception branch includes: a temporal memory network and a temporal context guidance module, wherein: The temporal memory network includes: feature connections, two-dimensional convolution blocks and multi-layer cascaded memory blocks; The temporal context guidance module includes: feature connection, channel rearrangement and two-dimensional convolution block.

[0011] Preferably, the processing process of the temporal memory network includes: Initial spatiotemporal features of the first frame image First, feature alignment is performed through a two-dimensional convolution block, and then the hidden state of the Lth layer is obtained through an L-layer memory block. and memory status ; Initial spatiotemporal features of other frame images ,in , first compared with the hidden state of the Lth layer of the previous frame Perform feature connection in the channel dimension, then align the features through a two-dimensional convolution block, and then update the hidden state of each layer memory block of the previous frame to obtain the updated hidden state of the Lth layer and memory status ; The hidden state of the Lth layer of each frame is obtained and the memory state of the Lth layer of the key frame The local time domain features of each frame image are obtained through a two-dimensional convolution block. and global temporal features of image sequences .

[0012] Preferably, the processing process of the time context guidance module includes: Local temporal features of each frame image and global temporal features of image sequences First, feature enhancement is performed through channel-position collaborative attention, followed by feature connection and channel rearrangement; The channel rearranged and enhanced features are processed through a series of two-dimensional convolution blocks to obtain features that are more time-sensitive. ,in ; Will Perform feature connection and obtain the time domain features output by the temporal evolution feature perception branch through two serially connected two-dimensional convolution blocks .

[0013] Preferably, the gate-based dual attention feature fusion module includes: channel-dimensional average pooling, channel-dimensional maximum pooling, global average pooling, feature connection, two-dimensional convolution block and parameter normalization; the specific processing process includes: Time domain characteristics Two parallel channel-dimensional average pooling and channel-dimensional maximum pooling are used to obtain two two-dimensional features, which are then connected and passed through a two-dimensional convolution block and a parameter normalization to obtain the spatial attention coefficient and the time domain feature. Weighted time domain features after selective feature enhancement ; Airspace characteristics First, it is processed by a global average pooling, and then two 2D convolution blocks in series and a parameter normalization are used to obtain the channel attention coefficient. Weighted spatial features are obtained after selective feature enhancement ; Enhanced time domain features and spatial characteristics The weighted coefficient is obtained through a feature connection, a series of two-dimensional convolution blocks and a series of parameter normalization. ,right and Complementary fusion is performed to obtain spatiotemporal correlation features through two serially connected two-dimensional convolution blocks .

[0014] Preferably, the target detection head includes: a two-dimensional convolution block, three parallel detection branches and a feature connection, wherein the first detection branch and the second detection branch are both composed of three series-connected two-dimensional convolution blocks, and the third detection branch is composed of a two-dimensional convolution block. The three detection branches respectively output the regression result, classification result and confidence result of the target, and the final target prediction box result is obtained after connecting them in the channel dimension through the feature connection.

[0015] Preferably, the method further comprises training a network for perceiving temporal evolution characteristics of small targets in dim space, wherein the overall loss function of the training is the sum of regression loss, confidence loss and classification loss, wherein: The regression loss adopts the intersection-over-union loss function; the confidence loss and classification loss both adopt the binary cross entropy loss function; Adam optimizer is used to update and predict network parameters, and the learning rate Set to 1e-4, training batch size is 4; The training includes the steps of preparing a training set, specifically including: Step A1: Set the observation position, observation start and end time, boresight direction, observation field angle, maximum detection distance, limit detection magnitude, camera aperture, exposure time, optical system quantum efficiency, optical transmittance, camera gain, camera grayscale bias, readout noise, dark current noise, image sensor target size, image resolution, image frame number, and target imaging signal-to-noise ratio simulation parameters of the space-based long-range passive optical observation platform; Step A2: Select the visible stars in the field of view according to the star catalog and all the parameters set in step A1, and use the position, proper motion, precession and other information in the star catalog to correct the star position and convert it into the imaging coordinate system; Step A3: Calculate the imaging grayscale values ​​of the visible stars in the field of view using the apparent magnitude information in the star catalog and simulation parameters, and use the Gaussian point spread function to generate star points with different diffusion degrees and widths. These points are added to the pure black background image according to the imaging coordinates to generate a background star map. Step A4: Establish a noise model for the imaging noise with known distribution, use zero-mean Gaussian white noise to model the imaging noise, generate imaging noise with different means and variances, and superimpose it on the background star image to generate a background image; Step A5: Calculate the positions of different targets with known orbital information, obtain the position information in the imaging coordinate system through coordinate transformation, and derive the target imaging grayscale value as follows based on the set target imaging signal-to-noise ratio value, Gaussian point spread function, and the mean and standard deviation of the generated background image: ; ; in, is the average gray value of a single target in the image, is the grayscale standard deviation of the generated background image, is the grayscale mean of the background image, is the target imaging signal-to-noise ratio set in the simulation parameters, is the imaging grayscale value at the center of the target, is the number of pixels occupied by the target, is the imaging coordinate at the center of the target, is the natural logarithm, The target imaging area k Points, The target imaging area k The imaging coordinates of the points; Step A6: Superimpose the grayscale value of the target area on the generated background image to obtain a target simulation image sequence, and superimpose the target in the pure black image as a label; Step A7: Construct the generated target simulation images and their labels into a dataset and divide them into a training set, a validation set, and a test set.

[0016] Compared with the prior art, the advantages of the present invention are: 1. In the initial feature extraction stage of the network, the present invention adopts a lightweight three-dimensional convolutional network to replace the traditional two-dimensional spatial feature extraction network. It can jointly extract the initial spatiotemporal features of the target from multiple consecutive frames of images, and use the timing information between frames to supplement the spatial representation of the target, effectively alleviating the problem of high similarity between the target and the stationary background stars in the spatial domain, thereby improving the distinguishability and detection performance of dim targets with low signal-to-noise ratio.

[0017] 2. In terms of spatiotemporal collaborative modeling, a dual-branch spatiotemporal structure is used to extract temporal and spatial target features. The temporal evolution feature perception branch models target motion trends by analyzing dynamic changes between consecutive frames. The spatial region enhancement branch, based on a cross-attention mechanism, leverages spatial difference information in the reference frame to enhance the target region in the current frame. This achieves spatially guided spatiotemporal feature collaborative enhancement, improving the accuracy and robustness of target positioning.

[0018] 3. To further explore the deep spatiotemporal correlation characteristics of the target, a gated dual-attention feature fusion module is adopted. The channel attention is used to selectively enhance the target responsiveness of the spatial domain features, and the spatial attention is used to highlight the target salient areas in the temporal domain features, further improving the discriminability of the feature representation. At the same time, a gating mechanism is introduced to fuse the enhanced features to achieve effective complementarity of spatiotemporal information and enhance the representation ability and detection performance of weak targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a schematic diagram of the network structure for perceiving the temporal evolution characteristics of small targets in dim space based on spatial domain guidance in the method of the present invention; Figure 2 It is a structural diagram of the temporal memory network, the gate-based dual attention feature fusion module and the target detection head in the method of the present invention. DETAILED DESCRIPTION

[0020] To solve these problems, the present invention proposes a method for detecting small dim spatial targets based on spatial guidance. The image sequence of the space target to be detected is input into a pre-established and trained spatial guidance-based dim spatial target temporal evolution feature perception network to obtain a target prediction frame, thereby realizing dim spatial target detection. Among them, the core invention is the spatial guidance-based dim spatial target temporal evolution feature perception network. The network first extracts the initial spatiotemporal features of the image sequence through a lightweight three-dimensional convolutional network to enhance the target response and suppress stellar clutter. Secondly, a dual-branch structure is constructed to model the target temporal evolution features and enhance the target spatial region features. Afterwards, a gated dual attention feature fusion module is used to selectively enhance and fuse the features output by the branch structure to improve the joint expression capability of position and motion information. Finally, end-to-end dim spatial target detection is achieved through the detection head.

[0021] In order to solve the problem of difficulty in extracting features of small targets in dark and weak space under complex deep space background, the present invention proposes a spatial domain-guided small target temporal evolution feature perception network, which includes the following steps: Step A: Establish an optical imaging model for the target observation image in the space-based long-distance passive optical detection scenario. Generate sequence image simulation data of space moving targets with different observation positions, different observation directions, different motion trajectories, and different signal-to-noise ratios through the optical imaging model to construct a simulation data set.

[0022] Step B: Construct a spatially guided temporal evolution feature perception network for dim and small spatial targets. This network consists of four modules: an initial feature extraction module, a spatiotemporal dual-branch feature extraction module, a gated dual-attention feature fusion module, and a target detection head. First, the initial spatiotemporal features of the target are extracted using the initial feature extraction module. Second, the spatiotemporal dual-branch feature extraction module extracts the target's temporal and spatial features, respectively. The spatiotemporal dual-branch feature extraction module consists of two branches: a temporal evolution feature perception branch and a cross-attention-based spatial region enhancement branch. Subsequently, the gated dual-attention feature fusion module fuses the target's temporal and spatial features obtained by the spatiotemporal dual-branch feature extraction module to obtain the target's spatiotemporal correlation features. Finally, the target detection head processes the target's spatiotemporal correlation features to obtain the final target detection result.

[0023] Step C: Set the loss function and training hyperparameters for the spatially guided dim and small target temporal evolution feature perception network, train the network using the simulation data set generated in step A, and obtain the optimal network parameter model.

[0024] Step D: Load the network parameter model and input the image sequence of the space target to be detected into the spatial domain guided small target temporal evolution feature perception network to obtain the target prediction frame and realize the detection of small target in dark space. The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0025] Example An embodiment of the present invention proposes a method for detecting small targets in dim space based on spatial guidance.

[0026] The implementation plan consists of four parts: first, a simulation dataset of space target sequence images in a space-based long-distance detection scenario is generated; then, a spatially guided dark and weak space small target temporal evolution feature perception network is constructed; then, the loss function and training hyperparameters are set, and the network is trained using the simulation dataset to obtain the optimal model; finally, the optimal model is loaded into the network, and the image sequence to be detected is input into the spatially guided dark and weak space small target temporal evolution feature perception network to obtain the target detection result.

[0027] Step 1: Create a dataset Step A: Establish an optical imaging model for the target observation image in the space-based long-distance passive optical detection scenario. Generate sequence image simulation data of space moving targets with different observation positions, different observation directions, different motion trajectories, and different signal-to-noise ratios through the optical imaging model to construct a simulation data set.

[0028] Step A1: Set the observation platform's observation position, observation start and end time, line of sight direction, observation field angle, maximum detection distance, limit detection magnitude, camera aperture, exposure time, optical system quantum efficiency, optical transmittance, camera gain, camera grayscale bias, readout noise, dark current noise, image sensor target size, image resolution, image frame number, and target imaging signal-to-noise ratio simulation parameters.

[0029] Step A2: Select the visible stars in the field of view according to the star catalog and all the parameters set in step A1, and use the position, proper motion, precession and other information in the star catalog to correct the star position and convert it into the imaging coordinate system.

[0030] Step A3: Use the apparent magnitude information in the star catalog and simulation parameters to calculate the imaging grayscale values ​​of the visible stars in the field of view, and use the Gaussian point spread function to generate star points with different diffusion degrees and widths. Add them to the pure black background image according to the imaging coordinates to generate a background star map.

[0031] Step A4: Establish a noise model for the imaging noise with known distribution. Use zero-mean Gaussian white noise to model the imaging noise. Generate imaging noise with different means and variances, and superimpose it on the background star image to generate a background image.

[0032] Step A5: Calculate the positions of different targets with known orbital information, obtain the position information in the imaging coordinate system through coordinate transformation, and derive the target imaging grayscale value based on the set target imaging signal-to-noise ratio value, Gaussian point spread function, and the mean and standard deviation of the generated background image. The derivation process is shown in formulas (1) and (2): ;(1) ;(2) in, is the average gray value of a single target in the image, is the grayscale standard deviation of the generated background image, is the grayscale mean of the background image, is the target imaging signal-to-noise ratio set in the simulation parameters, is the imaging grayscale value at the center of the target, is the number of pixels occupied by the target, is the imaging coordinate at the center of the target, is the natural logarithm, The target imaging area k Points, The target imaging area k The imaging coordinates of a point.

[0033] Step A6: Superimpose the grayscale value of the target area on the generated background image to obtain a target simulation image sequence, and superimpose the target in the pure black image as a label.

[0034] Step A7: The generated target simulation images and their labels are constructed into a dataset, which is further divided into a training set, a validation set, and a test set.

[0035] Step 2: Build a spatial domain-guided network for perceiving the temporal evolution characteristics of small targets in dim space. Step B: Construct a spatially guided temporal evolution feature perception network for dim and small spatial targets. This network consists of four modules: an initial feature extraction module, a spatiotemporal dual-branch feature extraction module, a gated dual-attention feature fusion module, and a target detection head. First, the initial spatiotemporal features of the target are extracted using the initial feature extraction module. Second, the spatiotemporal dual-branch feature extraction module extracts the target's temporal and spatial features, respectively. The spatiotemporal dual-branch feature extraction module consists of two branches: a temporal evolution feature perception branch and a cross-attention-based spatial region enhancement branch. Subsequently, the gated dual-attention feature fusion module fuses the target's temporal and spatial features obtained by the spatiotemporal dual-branch feature extraction module to obtain the target's spatiotemporal correlation features. Finally, the target detection head processes the target's spatiotemporal correlation features to obtain the final target detection result.

[0036] Step B1: The spatial-domain guided dim spatial small target temporal evolution feature perception network consists of four modules: initial feature extraction module, spatiotemporal dual-branch feature extraction module, gated dual attention feature fusion module and target detection head.

[0037] Step B2: In the initial feature extraction module, the input continuous 5-frame image sequence passes through the initial feature extraction module to obtain the initial spatiotemporal features of each frame. Figure 1 As shown in Figure 1, it contains a time dimension stacking operation, three 3D convolution blocks, three downsampling blocks and two 2D convolution blocks. The input continuous 5-frame image sequence is ,in, is the reference frame, The input image sequence is stacked in the time dimension to obtain the spatiotemporal tensor ( is the number of channels of the input image, is the number of image rows, is the number of image columns).

[0038] Step B3: After obtaining the space-time tensor Afterwards, three sets of alternating series-connected 3D convolutional blocks (C3BS) and downsampling (DownSample Conv3d—BatchNorm3d—SiLU, D3BS) are used to extract the target spatiotemporal features layer by layer, gradually reducing the spatial resolution of the features. The 3D convolutional block consists of a 3D convolutional layer, batch normalization, and a nonlinear activation function connected in sequence (Conv3d—BatchNorm3d—SiLU, C3BS). Downsampling uses a 3D convolutional layer with a stride of 2, batch normalization, and a nonlinear activation function connected in sequence (DownSample Conv3d—BatchNorm3d—SiLU, D3BS). The processing result is shown in Formula (3).

[0039] (3) in, is the target spatiotemporal feature obtained after the third downsampling.

[0040] Step B4: After obtaining the target spatiotemporal features After that, the time dimension feature is split to obtain the spatiotemporal features corresponding to each frame image. ( ), which is looped into two serial two-dimensional convolution blocks to obtain the initial spatiotemporal features of each frame image ( The 2D convolutional block consists of a 2D convolutional layer, batch normalization, and a nonlinear activation function connected in sequence (Conv2d—BatchNorm2d—SiLU, C2BS). The initial spatiotemporal feature extraction process is shown in Formula (4).

[0041] (4) Step B5: Obtain the initial spatiotemporal features of the input image sequence ( ), the target time domain and spatial domain features are extracted respectively through the spatiotemporal dual-branch feature extraction module. The spatiotemporal dual-branch feature extraction module consists of two branches, one is the temporal evolution feature perception branch, and the other is the spatial region enhancement branch based on cross attention. Figure 1 As shown in Figure 2. The spatial region enhancement branch based on cross attention mainly consists of feature connection, two-dimensional convolution block, cross attention and parameter normalization. The initial spatiotemporal features of the reference frame After inputting the cross-attention-based spatial region enhancement branch, the channel-dimensional feature concatenation (Feature Concatenation, Concat) is first performed, and the spatial difference features of the reference frame are further extracted through two serially connected two-dimensional convolution blocks. Subsequently, three parallel two-dimensional convolution blocks are used to generate the query (Query, Q) corresponding to the key frame, and the key (Key, K) and value (Value, V) features corresponding to the reference frame. Based on the calculation formula of the attention mechanism, the key frame spatial features weighted by the cross-attention coefficient are further obtained to achieve the enhancement of the target spatial region. The processing process is shown in formulas (5) to (8).

[0042] (5) (6) (7) (8) in, is the cross attention coefficient, is the Hadamard product operation (i.e. element-wise multiplication), It is the key frame spatial feature weighted by the cross attention coefficient.

[0043] Step B6: In order to make full use of the spatial difference information of the reference frame and retain the characteristics of the key frame itself, the initial spatiotemporal features of the key frame are converted to Keyframe spatial features weighted with cross-attention coefficients Perform residual connection. The residual connection features are then passed through three parallel two-dimensional convolution blocks to extract multi-scale features of the target area. The multi-scale features are summed and parameterized to obtain spatial weighting coefficients. The key frame features are weighted to enhance the response to target areas of different sizes and improve the discriminative ability of spatial features. The processing process is shown in formulas (9) and (10).

[0044] (9) (10) in, is the initial spatiotemporal features of the key frame Keyframe spatial features weighted with cross-attention coefficients The features obtained after residual connection, The kernel sizes are 、 and The two-dimensional convolution block, is the parameter normalization operation, The spatial features of the branch output are enhanced for the spatial regions based on the cross attention.

[0045] Step B7: At the same time, the temporal evolution feature perception branch extracts the temporal features of the target. The temporal evolution feature perception branch mainly consists of a temporal memory network and a temporal context guidance module, such as Figure 1 As shown in Figure 2, the temporal context guidance module consists of feature connections, channel shuffle (CS), and two-dimensional convolution blocks. The temporal memory network consists of feature connections, two-dimensional convolution blocks, and multi-layer cascaded memory blocks, as shown in Figure 2. Figure 2 The memory block is a spatiotemporal long short-term memory network structure (STLSTM). The initial spatiotemporal features output by the initial feature extraction module are First, input it into the temporal memory network in chronological order, and obtain the local time domain features of each frame image after passing through L layers of memory blocks. and global temporal features of image sequences The specific implementation process is: the initial spatiotemporal features of the first frame image First, feature alignment is performed through a two-dimensional convolution block, and then the hidden state of the Lth layer is obtained through an L-layer memory block. and memory status . The initial spatiotemporal features of other frames are then ( ) first with the hidden state of the Lth layer of the previous frame Perform feature connection in the channel dimension, then align the features through a two-dimensional convolution block, and then update the hidden state of each layer memory block of the previous frame to obtain the updated hidden state of the Lth layer and memory status Finally, the hidden state of the Lth layer of each frame is obtained and the memory state of the Lth layer of the key frame The local time domain features of each frame image are obtained through a two-dimensional convolution block. and global temporal features of image sequences The process is shown in formulas (11) to (16).

[0046] (11) (12) (13) (14) (15) (16) in, is the initial spatiotemporal features of the first frame image After feature alignment, For the Initial spatiotemporal characteristics of frame images ( ) and the hidden state of the Lth layer in the previous frame Features that are connected and then aligned. For the Frame image The cell state of the layer memory block, For the Frame image The hidden state of the layer memory block, For the Layer memory blocks.

[0047] Step B8: In the temporal context guidance module, the local temporal features of each frame ( ) and global temporal features of image sequences First, the network’s ability to represent features is enhanced by feature enhancement (using Channel-Position Collaborative Attention, CPCA). Then, the local temporal features of each frame are ( ) and global temporal features of image sequences Perform feature connection and enhance the information interaction between features by channel rearrangement to achieve guided enhancement of local time domain features of each frame. Afterwards, the channel rearranged features are processed by a series of two-dimensional convolution blocks to obtain features that are more time-sensitive. ( ). Finally, Perform feature connection and obtain the time domain features output by the temporal evolution feature perception branch through two series-connected two-dimensional convolution blocks The processing process is shown in formulas (17) and (18).

[0048] (17) (18) Step B9: The gated dual attention feature fusion module consists of channel-wise average pooling (CAP), channel-wise max pooling (CMP), global average pooling (GAP), feature concatenation, two-dimensional convolution blocks and parameter normalization, as shown in Figure 9. Figure 2As shown. Spatial features obtained by the spatiotemporal dual-branch structure and time domain characteristics Selective feature enhancement and fusion are performed through a gate-based dual attention feature fusion module. Among them, the temporal feature First, two parallel channel-dimensional average pooling and channel-dimensional maximum pooling are used to obtain two two-dimensional features. Then, the two two-dimensional features are connected, and then a two-dimensional convolution block and a parameter normalization are used to obtain the spatial attention coefficient, and the time domain features are normalized. Weighted time domain features after selective feature enhancement The processing process is shown in formulas (19) and (20).

[0049] (19) (20) in, is the spatial attention coefficient.

[0050] Step B10: Spatial Characterization First, it is processed by a global average pooling, then two 2D convolution blocks in series and a parameter normalization are used to obtain the channel attention coefficient, and finally the spatial domain features are processed. Weighted spatial features are obtained after selective feature enhancement The processing process is shown in formulas (21) and (22).

[0051] (twenty one) (twenty two) in, is the channel attention coefficient.

[0052] Step B11: Enhanced time domain features and spatial characteristics The weighted coefficient is obtained through a feature connection, a series of two-dimensional convolution blocks and a series of parameter normalization. , then the two features are complementary fused, and finally the spatiotemporal correlation features output by the gated dual attention feature fusion module are obtained through two series of two-dimensional convolution blocks. The processing process is shown in formulas (23) and (24).

[0053] (twenty three) (twenty four) Step B12: Spatiotemporal correlation features output by the gated dual attention feature fusion module Finally, it is sent to the target detection head to obtain the final target detection result. The structure of the target detection head is as follows Figure 2 As shown in the figure, it consists of a 2D convolutional block, three parallel detection branches, and a feature connection. The first and second detection branches each consist of three serially connected 2D convolutional blocks, while the third detection branch consists of a single 2D convolutional block. The three detection branches output the target regression result, classification result, and confidence result, respectively. The final target detection result is obtained by connecting them in the channel dimension through the feature connection.

[0054] Step 3: Network training Step C: Set the loss function and training hyperparameters for the spatially guided dim and small target temporal evolution feature perception network, train the network using the simulation data set generated in step A, and obtain the optimal network parameter model.

[0055] Step C1: For the spatial domain-guided dim spatial small target temporal evolution feature perception network, the overall loss function is the sum of regression loss, confidence loss, and classification loss. Among them, the regression loss uses the intersection-over-union (IoU) loss function to measure the degree of overlap between the predicted box and the true box; the confidence loss and classification loss both use the binary cross entropy (BCE) loss function to optimize the accuracy of target existence judgment and category prediction, respectively. In order to adapt to a single classification task, the hyperparameter Weight the classification loss to reduce the impact of the classification loss on the overall loss function. And use hyperparameters The regression loss is weighted to enhance the impact of the regression loss on the overall loss function, thereby improving the accuracy of the prediction box.

[0056] Step C2: Use Adam optimizer to update and predict network parameters, learning rate Set to 1e-4, and the training batch size is 4. Initialize the training parameters of the spatial domain-guided dim space small target temporal evolution feature perception network, load the training set and validation set, train the network, and obtain the optimal parameter model of the spatial domain-guided dim space small target temporal evolution feature perception network.

[0057] Step D: Load the network parameter model and input the space target image sequence to be detected into the spatial domain guided dim space small target temporal evolution feature perception network to obtain the detected target image sequence.

[0058] Step D1: Create a spatially guided dark and weak small target temporal evolution feature perception network and load the optimal parameter model trained in step C.

[0059] Step D2: Input the space target image sequence to be detected into the spatial domain-guided dim space small target temporal evolution feature perception network to obtain the target prediction box, and draw the target prediction box in the original target image to obtain the target detection result.

[0060] It is worth noting that in the embodiment of the above system, the modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of the functional modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of the present invention.

[0061] Summary: This invention provides a method for detecting faint small targets in space based on spatial guidance. It designs a spatially guided temporal evolution feature perception network for faint small targets. This method replaces traditional two-dimensional convolutional networks with a lightweight three-dimensional convolutional network to jointly extract the target's spatiotemporal features from the image sequence to be detected, effectively addressing issues such as weak target features and high similarity to stellar clutter. To model the dynamic evolution of targets between consecutive frames, the network introduces a temporal evolution feature perception branch. Simultaneously, a cross-attention-based spatial region enhancement branch is designed to utilize semantic information from the reference frame to guide the spatial localization of the target in the current frame, thereby addressing the insufficient spatial discrimination caused by independent processing of spatial and temporal features. To fully integrate multidimensional features, a gated dual-attention feature fusion module is further introduced to dynamically integrate the target's spatiotemporal information, enhancing the ability to express spatiotemporal coordination and thus achieving accurate detection of faint targets. Therefore, this invention can significantly improve the detection performance of faint small targets in low signal-to-noise ratio and complex deep space scenes, demonstrating promising engineering practicality and widespread application.

[0062] Finally, it should be noted that the above embodiments are intended only to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the embodiments, it should be understood by those skilled in the art that modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention and are intended to be encompassed by the claims of the present invention.

Claims

1. A method for detecting small targets in dim space based on spatial guidance, comprising: The image sequence of the space target to be detected is input into the pre-established and trained spatial domain-guided dim space small target temporal evolution feature perception network to obtain the target prediction frame and realize dim space small target detection; The proposed network for perceiving the temporal evolution of small targets in dim space extracts the initial spatiotemporal features of image sequences through a lightweight three-dimensional convolutional network, enhances target response and suppresses stellar clutter. By constructing a dual-branch structure, it models the temporal evolution features of targets and enhances the spatial regional features of targets. A gated dual-attention feature fusion module is used to selectively enhance and fuse the features output by the branch structure, thereby improving the joint expression capability of position and motion information.

2. The method for detecting small targets in dim space based on spatial guidance according to claim 1, characterized in that: The input of the dim space small target temporal evolution feature perception network is a continuous 5-frame image sequence, and the output is the target prediction frame of the last frame of the current image sequence. The dim space small target temporal evolution feature perception network includes: an initial feature extraction module, a spatiotemporal dual-branch feature extraction module, a gated dual attention feature fusion module and a target detection head, wherein, The initial feature extraction module is used to extract 5 consecutive frames of images of the space target to be detected. Perform initial spatiotemporal feature extraction to obtain the corresponding initial spatiotemporal features ; The spatiotemporal dual-branch feature extraction module is used to extract the target time domain and spatial domain features respectively through the temporal evolution feature perception branch and the spatial region enhancement branch based on cross attention, and obtain the time domain features. and spatial characteristics ; The gate-based dual attention feature fusion module is used to fusion the temporal features and spatial characteristics Fusion is performed to obtain the target spatiotemporal correlation features ; The target detection head is used to detect the temporal and spatial correlation features of the target Processing is performed to obtain the final target prediction box result.

3. The method for detecting small targets in dim space based on spatial guidance according to claim 2, characterized in that: The initial feature extraction module includes: a sequentially connected time dimension stack, three groups of alternating series-connected three-dimensional convolution blocks and downsampling, and two two-dimensional convolution blocks; wherein the three-dimensional convolution block includes a sequentially connected three-dimensional convolution layer, batch normalization, and a nonlinear activation function; the downsampling includes a sequentially connected three-dimensional convolution layer with a stride of 2, batch normalization, and a nonlinear activation function, and the two-dimensional convolution block includes a sequentially connected two-dimensional convolution layer, batch normalization, and a nonlinear activation function; The specific processing process includes: Input a sequence of 5 consecutive image frames ,in, is the reference frame, is a key frame, and the space-time tensor is obtained by stacking in the time dimension ,in, is the number of channels of the input image, is the number of image rows, is the number of image columns; After three sets of alternating series of three-dimensional convolution blocks and downsampling, the target spatiotemporal features are obtained. , perform feature splitting in the time dimension and obtain the i Spatiotemporal features corresponding to frame images ,in ; Will The loop inputs two 2D convolution blocks in series and obtains the i Initial spatiotemporal characteristics of frame images .

4. The method for detecting small targets in dim space based on spatial guidance according to claim 2, characterized in that: The spatial region enhancement branch based on cross attention includes: feature connection, two serial two-dimensional convolution blocks, three parallel two-dimensional convolution blocks, cross attention, two-dimensional convolution block, three parallel two-dimensional convolution blocks and parameter normalization. The specific processing process includes: Initial spatiotemporal characteristics of the reference frame First, the channel dimension feature connection is performed, and the spatial difference features of the reference frame are further extracted through two serial two-dimensional convolution blocks. The query Q corresponding to the key frame, the key K and value V corresponding to the reference frame are generated through three parallel two-dimensional convolution blocks respectively; based on the attention mechanism, the key frame spatial features weighted by the cross attention coefficient are obtained ; The initial spatiotemporal features of the keyframes and Perform residual connection, and extract the multi-scale features of the target area through three parallel two-dimensional convolution blocks. Then, add the multi-scale features and obtain the spatial weighting coefficient after parameter normalization. Then, weight the key frame features and output the spatial domain features. .

5. The method for detecting small targets in dim space based on spatial guidance according to claim 2, characterized in that: The temporal evolution feature perception branch includes: a temporal memory network and a temporal context guidance module, wherein: The temporal memory network includes: feature connections, two-dimensional convolution blocks and multi-layer cascaded memory blocks; The temporal context guidance module includes: feature connection, channel rearrangement and two-dimensional convolution block.

6. The method for detecting small targets in dim space based on spatial guidance according to claim 5, characterized in that: The processing process of the temporal memory network includes: Initial spatiotemporal features of the first frame image First, feature alignment is performed through a two-dimensional convolution block, and then the hidden state of the Lth layer is obtained through an L-layer memory block. and memory status ; Initial spatiotemporal features of other frame images ,in , first compared with the hidden state of the Lth layer of the previous frame Perform feature connection in the channel dimension, then align the features through a two-dimensional convolution block, and then update the hidden state of each layer memory block of the previous frame to obtain the updated hidden state of the Lth layer and memory status ; The hidden state of the Lth layer of each frame is obtained and the memory state of the Lth layer of the key frame The local time domain features of each frame image are obtained through a two-dimensional convolution block. and global temporal features of image sequences .

7. The method for detecting small targets in dim space based on spatial guidance according to claim 6, characterized in that: The processing process of the time context guidance module includes: Local temporal features of each frame image and global temporal features of image sequences First, feature enhancement is performed through channel-position collaborative attention, followed by feature connection and channel rearrangement; The channel rearranged and enhanced features are processed through a series of two-dimensional convolution blocks to obtain features that are more time-sensitive. ,in ; Will Perform feature connection and obtain the time domain features output by the temporal evolution feature perception branch through two serially connected two-dimensional convolution blocks .

8. The method for detecting small targets in dim space based on spatial guidance according to claim 2, characterized in that: The gated dual attention feature fusion module includes: channel-dimensional average pooling, channel-dimensional maximum pooling, global average pooling, feature connection, two-dimensional convolution block and parameter normalization; the specific processing process includes: Time domain characteristics Two parallel channel-dimensional average pooling and channel-dimensional maximum pooling are used to obtain two two-dimensional features, which are then connected and passed through a two-dimensional convolution block and a parameter normalization to obtain the spatial attention coefficient and the time domain feature. Weighted time domain features after selective feature enhancement ; Airspace characteristics First, it is processed by a global average pooling, and then two 2D convolution blocks in series and a parameter normalization are used to obtain the channel attention coefficient. Weighted spatial features are obtained after selective feature enhancement ; Enhanced time domain features and spatial characteristics The weighted coefficient is obtained through a feature connection, a series of two-dimensional convolution blocks and a series of parameter normalization. ,right and Complementary fusion is performed to obtain spatiotemporal correlation features through two serially connected two-dimensional convolution blocks .

9. The method for detecting small targets in dim space based on spatial guidance according to claim 2, characterized in that: The target detection head includes: a two-dimensional convolution block, three parallel detection branches and a feature connection, wherein the first detection branch and the second detection branch are each composed of three series-connected two-dimensional convolution blocks, and the third detection branch is composed of one two-dimensional convolution block. The three detection branches respectively output the regression result, classification result and confidence result of the target, and the final target prediction box result is obtained after being connected in the channel dimension through the feature connection.

10. The method for detecting small targets in dim space based on spatial guidance according to claim 1, characterized in that: The method also includes training a network for perceiving small targets in dim and weak space with temporal evolution characteristics. The overall loss function of the training is the sum of regression loss, confidence loss, and classification loss, wherein: The regression loss adopts the intersection-over-union loss function; the confidence loss and classification loss both adopt the binary cross entropy loss function; Adam optimizer is used to update and predict network parameters, and the learning rate Set to 1e-4, training batch size is 4; The training includes the steps of preparing a training set, specifically including: Step A1: Set the observation position, observation start and end time, boresight direction, observation field angle, maximum detection distance, limit detection magnitude, camera aperture, exposure time, optical system quantum efficiency, optical transmittance, camera gain, camera grayscale bias, readout noise, dark current noise, image sensor target size, image resolution, image frame number, and target imaging signal-to-noise ratio simulation parameters of the space-based long-range passive optical observation platform; Step A2: Select the visible stars in the field of view according to the star catalog and all the parameters set in step A1, and use the position, proper motion, precession and other information in the star catalog to correct the star position and convert it into the imaging coordinate system; Step A3: Calculate the imaging grayscale values ​​of the visible stars in the field of view using the apparent magnitude information in the star catalog and simulation parameters, and use the Gaussian point spread function to generate star points with different diffusion degrees and widths. These points are added to the pure black background image according to the imaging coordinates to generate a background star map. Step A4: Establish a noise model for the imaging noise with known distribution, use zero-mean Gaussian white noise to model the imaging noise, generate imaging noise with different means and variances, and superimpose it on the background star image to generate a background image; Step A5: Calculate the positions of different targets with known orbital information, obtain the position information in the imaging coordinate system through coordinate transformation, and derive the target imaging grayscale value as follows based on the set target imaging signal-to-noise ratio value, Gaussian point spread function, and the mean and standard deviation of the generated background image: ； ； in, is the average gray value of a single target in the image, is the grayscale standard deviation of the generated background image, is the grayscale mean of the background image, is the target imaging signal-to-noise ratio set in the simulation parameters, is the imaging grayscale value at the center of the target, is the number of pixels occupied by the target, is the imaging coordinate at the center of the target, is the natural logarithm, The target imaging area k Points, The target imaging area k The imaging coordinates of the points; Step A6: Superimpose the grayscale value of the target area on the generated background image to obtain a target simulation image sequence, and superimpose the target in the pure black image as a label; Step A7: Construct the generated target simulation images and their labels into a dataset and divide them into a training set, a validation set, and a test set.

Images