A natural language motion prior guided spatial dim target detection method

Abstract

The application discloses a natural language motion prior guided spatial dim and weak target detection method, and belongs to the field of spatial target detection and situation awareness. The method designs a structured natural language description for a moving target, encodes the quadrant direction, distance area, motion speed and direction of the target between continuous frames; a visual-text motion feature alignment module is constructed based on a variational autoencoder, cross-modal alignment is realized in a hidden space, motion prior knowledge is learned by a visual branch; in the training stage, motion prior is implicitly transferred to the visual embedding module through back propagation, so that the detection head is prevented from being polluted by motion noise; in the reasoning stage, no text input is needed, and only multiple optical images can be used to output the target position, category and motion state in an end-to-end mode. Experiments show that under the condition of 2 frames of extremely short sequence, all indexes are relatively improved by about 4.5%, and under the condition of 2 to 5 frames, the method is better than a pure visual method, and can provide technical support for the safety of a spacecraft in orbit and spatial situation awareness.

Description

Technical Field

[0001] This invention belongs to the field of space target detection and space situational awareness technology, specifically relating to a method and device for detecting faint targets in space guided by natural language motion priors. Background Technology

[0002] With the increasing frequency of human space exploration activities, the Earth orbit space environment is becoming unprecedentedly crowded, the amount of space debris is growing exponentially, and the risk of space collisions is rising sharply. Space Situational Awareness (SSA) has risen to the level of a core requirement for maintaining national space security and ensuring the stable operation of spacecraft. Among various space target monitoring methods, ground-based optical observation has become the main means of long-term space target observation due to its advantages such as simple installation, no geographical limitations, and readily available measurement data. However, ground-based optical observation faces bottlenecks such as weak and dense targets, insufficient accuracy in orbital parameter estimation, and poor real-time situation assessment. Especially under the condition of sparse data in short frames and short arcs, the detection and identification of weak and faint space targets is extremely difficult.

[0003] Traditional methods for detecting faint targets in space primarily rely on "Detection by Tracking" (DBT) and "Tracking by Detection" (TBD) techniques based on classical digital image processing. The DBT algorithm preprocesses a single frame image, performs background suppression and star image extraction, and then uses trajectory association techniques to match suspected targets. While computationally fast, its accuracy is limited. The TBD algorithm, on the other hand, constructs an integrated detection and tracking process, accumulating and enhancing the signal-to-noise ratio of potential targets based on multiple frames of images without prior knowledge of the target's optical characteristics. Although it achieves higher accuracy, its computational speed is slower. These traditional methods generally suffer from poor robustness of manually designed features, weak generalization ability, and difficulty adapting to complex background changes.

[0004] In recent years, with the development of deep learning technology, object detection methods based on convolutional neural networks and Transformers have made significant progress. Convolutional neural network detection models, such as YOLO and Faster R-CNN, and Transformer detection models, such as DETR, have achieved automated feature extraction and object localization. Subsequently, addressing the characteristics of weak and small objects in spatial object detection tasks, researchers have further proposed spatial object detection methods such as CSAUNet, DnTNet, and SENet, achieving some progress in weak feature enhancement and object detection.

[0005] However, faint targets in space often have a high similarity to complex backgrounds, and the target's morphological information is not obvious, making it difficult to accurately distinguish between the target and the background. In addition, since the motion of faint targets in space is extremely weak, the displacement difference between adjacent frames of optical images is small. After multiple downsampling by deep networks, the network is unable to effectively capture this weak motion information. As a result, existing spatial target detection algorithms are prone to losing faint target features during feature extraction, and the detection accuracy is still limited. Summary of the Invention

[0006] To address the aforementioned technical challenges, this invention proposes a natural language motion prior-guided method for detecting dimly lit targets in space. By constructing a structured natural language description of the target's motion state and establishing a visual-text motion feature alignment model, motion prior knowledge, which is difficult to learn directly from visual observation, is transformed into learnable natural language semantic representations, achieving implicit supervision of visual feature learning by motion priors. During the training phase, the natural language description guides the visual embedding module to learn the target's motion patterns and spatiotemporal features, improving the model's ability to perceive and represent dimly lit targets. During the inference phase, no additional text input is required; high-precision detection of dimly lit spatial targets can be achieved solely through multiple frames of optical observation images.

[0007] To achieve the above objectives, the present invention proposes the following technical solution:

[0008] A method for detecting faint targets in space guided by natural language motion priors, the method comprising:

[0009] Step 1: Design a natural language description for moving targets: Establish a polar coordinate system with the center of the observed image as the origin, divide the 360° space into multiple quadrants, divide the distance levels according to the distance between the target pixel and the center of the image, divide the speed levels according to the Euclidean distance between the target pixel and the two frames of images, divide the motion direction according to the direction of the displacement vector, and generate a structured natural language description containing the core elements of target number, quadrant orientation, distance region, motion speed and motion direction.

[0010] Step 2: Establish a latent space visual-text motion learning model: Construct a visual-text motion feature alignment module based on a variational autoencoder. The module includes a text branch and a visual branch. The text branch takes the structured natural language description generated in Step 1 as input, obtains the corresponding text embedding features through the text embedding module, estimates the variational posterior distribution of the text latent variables through the text encoder, and reconstructs the text motion prior feature map using the text decoder. The visual branch takes multi-frame spatial target observation images as input, obtains the corresponding visual embedding features through the visual embedding module, estimates the variational posterior distribution of the visual latent variables through the visual encoder, and reconstructs the visual motion prior feature map that is spatially aligned with the text motion prior feature map using the visual decoder. By constraining the consistency between the text motion prior and the visual motion prior, and introducing latent variable consistency loss and bidirectional contrastive loss based on InfoNCE, cross-modal alignment between the text latent space and the visual latent space is achieved. This allows the visual branch to learn the motion prior knowledge contained in the structured natural language description during training, thereby generating an equivalent motion prior representation without the participation of the text branch during the inference stage.

[0011] Step 3: Detection of spatially dim targets based on implicit motion prior supervision. During the training phase, the visual-text motion feature alignment module implicitly transfers motion prior knowledge from the text description to the visual embedding module through shallow, mid-, and deep consistency losses and bidirectional contrast losses. The multi-frame visual embedding features output by the visual embedding module gradually align with the text motion description during the learning process, thus internalizing the target's motion information. During the inference phase, the visual embedding features output by the visual embedding module alone possess sufficient motion perception capabilities and are directly fed into the detection head, which outputs the target's position, category, and motion state information. The visual-text motion feature alignment module is only used during the training phase; the visual branch runs independently during the inference phase without requiring additional text input.

[0012] On the other hand, the present invention provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned natural language motion prior-guided spatial faint target detection method.

[0013] Thirdly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned method for detecting spatially faint targets guided by natural language motion priors.

[0014] The beneficial effects of this invention are as follows:

[0015] (1) Significantly improves detection accuracy under extremely short frame conditions. Experiments show that under extremely short sequence conditions with only 2 frames of images as input, the F1 score of the method of the present invention reaches 0.9290, which is more than 4.5 percentage points higher than the 0.8886 of the pure vision method, effectively alleviating the performance degradation problem under extreme observation conditions.

[0016] (2) The detection performance is almost always better than the pure vision method under the conditions of 2 to 5 frames. Experimental verification shows that the F1 scores of the method of the present invention are 0.9290, 0.9362, 0.9402 and 0.9613 under the conditions of 2 frames, 3 frames, 4 frames and 5 frames respectively, which are significantly better than the pure vision method without motion prior (corresponding F1 scores of 0.8886, 0.9246, 0.9301 and 0.9466 respectively), verifying the universal advantage of the natural language motion prior guided target detection framework in the task of detecting targets in dim light. Attached Figure Description

[0017] Figure 1 Generate schematic diagrams for homogeneous natural language descriptions of moving targets;

[0018] Figure 2 A diagram illustrating a spatially faint target detection method guided by natural language motion priors.

[0019] Figure 3 This is a comparison chart of the detection results of the present invention and four advanced methods. The red boxes represent predicted locations, and the blue boxes represent true location locations. Detailed Implementation

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] like Figure 1 As shown, this invention proposes a spatial faint target detection method guided by natural language motion priors, which consists of three steps: constructing a homogeneous natural language description for moving targets, establishing a motion learning model with a shared visual-text latent space, and implicit motion prior supervision and target detection. Specifically, traditional ground-based optical observations struggle to effectively capture weak motion information under sparse data conditions of short frames and short arcs, leading to a significant decrease in the accuracy of faint target detection; existing deep learning methods lack prior constraints on the physical motion laws of targets, making it difficult to achieve fine motion feature modeling under complex stellar backgrounds and strong noise interference; general visual language models are difficult to directly adapt to spatial target detection scenarios, and the lack of text modality in the inference stage renders multimodal alignment schemes ineffective. This invention achieves high-precision recognition of spatial faint targets under short frame and short arc conditions by embedding motion text descriptions in the training stage and efficient detection using pure vision in the inference stage. The method specifically includes:

[0022] Step 1: Design homogeneous natural language descriptions for moving targets: such as Figure 1 As shown, this description system is generated based on the information of the target in a single frame and the motion state between two consecutive frames, and includes five core elements: target number, quadrant orientation, distance region, motion speed, and motion direction.

[0023] First of all, with Pixel Image Center A polar coordinate system is established with the origin, dividing the 360° space into 12 quadrants, each spanning 30°, to describe the target's orientation relative to the image center. The quadrants are numbered clockwise from due east, starting from A to L.

[0024] Secondly, based on the distance between the target pixel and the image center, the distance region is divided into five levels: center, core, middle, perimeter, and edge. Here, center indicates the target is located within a very small area in the center of the image; core indicates the target is close to the center; middle indicates the target is in the middle of the image; perimeter indicates the target is near the image edge; and edge indicates the target is at the outermost edge of the image.

[0025] Furthermore, based on the Euclidean distance between target pixels in two frames, the motion speed is divided into four levels: stationary, slow, medium, and fast. The specific thresholds can be determined based on the statistical distribution of target motion speed in the dataset. Pixel displacements of less than 2 pixels are defined as stationary, 2 to 5 pixels as slow, 5 to 10 pixels as medium, and greater than 10 pixels as fast.

[0026] Finally, based on the components of the displacement vector on the horizontal and vertical axes, the motion direction is divided into eight directions: east, northeast, north, northwest, west, southwest, south, and southeast, with each direction spanning 45°.

[0027] Thus, a complete natural language description example is: "Target 1 is located in quadrant J, near the perimeter, and is moving slowly toward the northeast." This description system accurately depicts the instantaneous motion state of the target in structured text form, providing high-quality semantic supervision signals for subsequent cross-modal alignment. During the training phase, for each target in each set of continuously observed images, a corresponding homogeneous natural language description is automatically generated. The text is then converted into fixed-dimensional text embedding features through a pre-trained text embedding module (with fixed parameters and not involved in training), thereby providing semantic prior constraints for latent space motion representation learning.

[0028] Step 2: Establish a motion learning model with shared visual-text latent space: such as Figure 2 As shown, this invention constructs a visual-text motion feature alignment module based on a variational autoencoder architecture, which includes text branches and visual branches, and achieves cross-modal knowledge transfer through latent space alignment and motion prior reconstruction.

[0029] In the text branch, the generated natural language description is used as input, and the text embedding module obtains the embedded features. Then, it is encoded using a text encoder to generate latent variables. variational posterior distribution Specifically, the text encoder outputs the mean of the latent variables. and variance Sampling is performed using reparameterization techniques:

[0030]

[0031] In the formula, ⊙ denotes element-wise multiplication, and ε is a noise vector sampled from a standard normal distribution. Subsequently, the text decoder will... Reconstructing a priori map of motion The diagram visually reflects the intensity and direction of the target's motion.

[0032] In the vision branch, multi-frame spatial target observation images are used as input, and visual embedding features are extracted by the visual embedding module (including a feature extraction skeleton network and a temporal feature aggregation layer). The latent variables are then encoded using a visual encoder. And use a visual decoder to reconstruct and generate visual motion priors. To achieve cross-modal alignment, a latent variable consistency loss is introduced to ensure that textual and visual latent variables are distributed consistently in the latent space.

[0033] Meanwhile, to prevent collapse during feature alignment, a bidirectional contrastive loss based on InfoNCE is introduced. Specifically, visual latent variables are... and textual implicit variables By mapping the image onto the contrast space using the projection head, we obtain... and Calculate their similarity By using symmetrical bidirectional contrastive loss, matching image-text pairs are brought closer together, while mismatched samples are pushed away.

[0034] In summary, the total loss of the visual-text motion feature alignment module is:

[0035] According to claim 1, the method for detecting weak spatial targets guided by natural language motion priors is characterized in that the total loss function of the visual-text motion feature alignment module in step 2 is:

[0036]

[0037] In the formula, Motion feature consistency loss is used to enable the visual model to learn visual embedding features that are implicitly aligned with the motion description in natural language, thereby enhancing the ability to perceive local motion. The target detection loss is composed of the basic detection loss; The variational autoencoder reconstruction loss is the reconstruction error generated by the decoder after sampling the latent variables generated by the text encoder using the reparameterization technique to reconstruct the motion prior feature map; α and β are balancing hyperparameters, which are set to 1 and 0.1 respectively in this invention.

[0038] More specifically, the motion feature consistency loss Furthermore, it is composed of the consistency loss of visual language motion features at three levels: shallow, middle, and deep. The calculation formula is as follows:

[0039]

[0040] In the formula, , and These are the visual language motion feature consistency losses for shallow, mid, and deep layers, with shallow consistency loss being the most significant. Constrained text embedding features With visual embedding features Alignment at the feature level; mid-level consistency loss text hidden variables Visual latent variables Distributional consistency constraints and InfoNCE-based bidirectional contrast constraints; deep consistency loss. Motion prior graph reconstructed for constrained text branches Motion priors generated with visual branching The mean square error between them; γ and θ are hyperparameters, which are set to 0.5 and 1 respectively in this invention.

[0041] The target detection loss The calculation formula is:

[0042]

[0043] In the formula, The number of positive samples. For CIoU loss, For target confidence loss, For classification loss.

[0044] The reconstruction loss The calculation formula is:

[0045]

[0046] In the formula, The real motion prior feature map is constructed from a heat map generated by simulation of ground truth values ​​of the sequence spatial target images; For textual implicit variables The motion prior feature map obtained by decoding and reconstruction; For visual latent variables The motion prior feature map obtained by decoding and reconstruction.

[0047] Compared to single-frame methods, multi-frame methods can extract additional motion features, enhancing the model's ability to detect small, faint, and dark targets. This step transfers prior motion knowledge from the language description to the visual branch, enabling the visual embedding module to learn the ability to perceive fine motion patterns of targets during the training phase without relying on any additional text input during the inference phase.

[0048] Step 3: Detection of weak targets in space based on implicit motion prior supervision: To address the problems of cluttered backgrounds in space images, varying target motion scales, and feature fusion noise contaminating the detection head, this invention designs an implicit motion prior supervision mode.

[0049] During the training phase, the visual-text motion feature alignment module, through backpropagation optimization using shallow, mid-level, and deep consistency loss and bidirectional contrast loss, directly injects prior knowledge of motion into the parameters of the visual embedding module. The multi-frame visual embedding features output by the visual encoder continuously align with the text motion description during the learning process, thereby internalizing the target's motion pattern, effectively suppressing background noise, and highlighting the target's motion region. In the inference phase, the visual embedding features output by the visual embedding module, already containing sufficient motion information, are directly fed into the detection head to output the target's position.

[0050] The detection head adopts a decoupled head structure based on YOLO, which includes a classification branch, a regression branch, and an IoU prediction branch. The classification branch is responsible for predicting the class confidence of the target, the regression branch is responsible for predicting the coordinate offset of the bounding box, and the IoU prediction branch is responsible for predicting the degree of overlap between the predicted box and the ground truth box.

[0051] During the training phase, the model simultaneously receives multiple frames of optical images and corresponding natural language descriptions of the moving targets. The visual branch learns prior motion knowledge by constraining the visual branch through the visual-text motion feature alignment module in step 2. During the inference phase, the visual-text motion feature alignment module is completely decoupled; only the multiple frames of optical images are input to the visual embedding module. Subsequently, the motion-aware visual embedding features output by the visual embedding module are fed into the detection head, and finally, the target's position, category, and motion state information are output. The entire process requires no additional text input, achieving the design goal of utilizing language priors during the training phase and performing efficient detection purely visually during the inference phase.

[0052] To verify the effectiveness of the method of this invention, experiments were conducted on a self-built multimodal dataset of spatial targets. This dataset was constructed through independent simulation and covers 100 sequences of four typical scenarios: low-Earth orbit, medium-Earth orbit, high-Earth orbit, and hybrid orbit. Each sequence contains 260 sets of optical images and YOLO format annotations. The experiments compared the detection performance of pure vision methods without motion description with that of the method of this invention under different frame rate conditions. The specific results are shown in the table below.

[0053] Table 1. Performance comparison between the proposed method and the pure vision method under different input frame numbers (optimal values ​​are in bold).

[0054]

[0055] As shown in Table 1, under the extremely short frame condition of 2 frames, the F1 score of the method of this invention reaches 0.9290, which is approximately 4.5% higher than the 0.8886 of the pure vision method, significantly alleviating the performance degradation problem under short frame conditions. Under conditions of 2 to 5 frames, the detection performance of the method of this invention is almost always better than the pure vision method without motion prior, verifying the universality advantage of this invention under various frame count conditions.

[0056] also, Figure 3 The detection results of the method of this invention are compared with those of advanced methods such as SSTNet, YOLOv8, and STME. Table 2 presents the corresponding quantitative evaluation results. Both subjective and objective experimental results show that the method of this invention outperforms the comparative methods in terms of detection accuracy. Furthermore, while maintaining high-precision detection performance, the method of this invention eliminates the need for a text encoder in the inference stage, reducing model computational complexity and meeting the requirements of real-time detection applications.

[0057] Table 2 Performance comparison of the method of this invention with four advanced methods (optimal values ​​are bolded)

[0058]

[0059] On the other hand, the present invention provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned natural language motion prior-guided spatial faint target detection method.

[0060] Thirdly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned method for detecting spatially faint targets guided by natural language motion priors.

[0061] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting weakly lit targets in space guided by natural language motion priors, characterized in that, The method includes: Step 1: Design a natural language (text) description for moving targets: Establish a polar coordinate system with the center of the observed image as the origin, divide the 360° space into multiple quadrants, divide the distance levels according to the distance between the target pixel and the image center, divide the velocity levels according to the Euclidean distance between the target pixel and the two frames, divide the motion direction according to the displacement vector direction, and generate a structured natural language description containing the core elements of target number, quadrant orientation, distance region, motion speed and motion direction; Step 2: Establish a latent space visual-text motion learning model: Construct a visual-text motion feature alignment module based on a variational autoencoder. This module includes a text branch and a visual branch. The text branch takes the structured natural language description generated in Step 1 as input, extracts text embedding features through the text embedding module, estimates the posterior distribution of text latent variables through the text encoder, and then reconstructs and generates a text motion prior feature map using the text decoder. The visual branch takes multi-frame spatial target observation images as input, extracts visual embedding features through the visual embedding module, estimates the posterior distribution of visual latent variables through the visual encoder, and then reconstructs and generates a visual motion prior feature map using the visual decoder. By constraining the consistency between text motion priors and visual motion priors, and combining latent variable consistency loss and bidirectional contrastive loss based on InfoNCE, cross-modal alignment between the text latent space and the visual latent space is achieved. This enables the visual branch to learn the target motion prior knowledge contained in the structured natural language description, thereby obtaining a visual representation with motion perception capabilities. Step 3: Implementing spatially faint target detection based on implicit motion prior supervision: During the training phase, the visual embedding module is implicitly supervised by the visual-text motion feature alignment module. Utilizing shallow, mid, and deep motion prior consistency constraints and bidirectional contrast loss, motion prior knowledge from structured natural language descriptions is transferred to visual embedding features. This allows the features extracted by the visual embedding module to gradually acquire target motion perception capabilities and effectively suppress stellar background noise while highlighting the target's motion region. During the inference phase, only the visual branch is retained. The visual embedding module directly outputs visual embedding features containing motion prior information and inputs them into the detection head to predict target location, target category, and motion state information. The visual-text motion feature alignment module is only used during the training phase; during the inference phase, the visual branch runs independently without additional text input.

2. The method for detecting weak spatial targets guided by natural language motion priors according to claim 1, characterized in that, The structured natural language description in step 1 is in the form of: "Target [number] is located in Quadrant [near the [distance region], is moving [speed level] toward the [direction]". Here, the quadrant divides the circumference into 12 sectors A, B, ..., L with a span of 30°. The distance region is divided into five levels: center, core, middle, perimeter, and edge, based on the distance between the target pixel and the image center. The speed level is divided into four levels: stationary, slow, moderate, and quick, based on the Euclidean distance between the target pixel and two frames. The direction of movement is divided into eight directions: east, northeast, north, northwest, west, southwest, south, and southeast.

3. The method for detecting weak spatial targets guided by natural language motion priors according to claim 1, characterized in that, The visual-text motion feature alignment module in step 2 includes a text branch and a visual branch; wherein, the text branch takes the structured natural language description generated in step 1 as input, and obtains text embedding features through the text embedding module. And estimate textual latent variables through a text encoder. The posterior distribution; the visual branch takes multiple frames of spatial target observation images as input, and obtains visual embedding features through the visual embedding module. And estimate visual latent variables through a visual encoder. The posterior distribution; the text decoder and visual decoder are based on the latent variables respectively. and Reconstruction yields prior textual motion. and visual-motor priors By aligning textual motion priors with visual motion priors, cross-modal mapping between textual and visual latent spaces is achieved, enabling the visual branch to acquire motion prior knowledge embedded in the natural language description. The joint probability distribution decomposes into the following steps for this visual-textual motion feature alignment process:

4. The method for detecting weak targets in space guided by natural language motion priors according to claim 1, characterized in that, The text branch reconstruction process in step 2 includes: assuming the reconstruction distribution is a multivariate Gaussian distribution, using the motion prior map generated by the decoder. As the mean, the reconstructed motion prior is approximated to the true motion prior by minimizing the text reconstruction loss; the text encoder, based on text embedding features... Adjusting the distribution of the latent space to make the latent variables It can effectively capture and express motion information in text. Therefore, visual latent variables... It is modulated and can provide sufficient information for the generation of its subsequent motion representation.

5. The method for detecting weak targets in space guided by natural language motion priors according to claim 1, characterized in that, In step 3, an implicit motion prior supervision method is adopted: during the training phase, the prior knowledge in the motion description text is transferred to the visual embedding module through the visual-text motion feature alignment module, so that the extracted visual embedding features themselves have motion perception capabilities, which are used to suppress stellar background noise and highlight the target motion area.

6. The method for detecting weak spatial targets guided by natural language motion priors according to claim 1, characterized in that, The total loss function of the visual-text motion feature alignment module in step 2 is: In the formula, Motion feature consistency loss is used to enable the visual model to learn visual embedding features that are implicitly aligned with the motion description in natural language, thereby enhancing the ability to perceive local motion. The target detection loss is composed of the basic detection loss; The variational autoencoder reconstruction loss is the reconstruction error generated by the decoder after sampling the latent variables generated by the text encoder using the reparameterization technique to reconstruct the motion prior feature map; α and β are balancing hyperparameters, which are set to 1 and 0.1 respectively in this invention. More specifically, the motion feature consistency loss Furthermore, it is composed of the consistency loss of visual language motion features at three levels: shallow, middle, and deep. The calculation formula is as follows: In the formula, , and These are the visual language motion feature consistency losses for shallow, mid, and deep layers, with shallow consistency loss being the most significant. Constrained text embedding features With visual embedding features Alignment at the feature level; mid-level consistency loss text hidden variables Visual latent variables Distributional consistency constraints and InfoNCE-based bidirectional contrast constraints; deep consistency loss. Motion prior graph reconstructed for constrained text branches Motion priors generated with visual branching The mean square error between them; γ and θ are hyperparameters, which are set to 0.5 and 1 respectively in this invention. The target detection loss The calculation formula is: In the formula, The number of positive samples. For CIoU loss, For target confidence loss, For classification loss. The reconstruction loss The calculation formula is: In the formula, The real motion prior feature map is constructed from a heat map generated by simulation of ground truth values ​​of the sequence spatial target images; For textual implicit variables The motion prior feature map obtained by decoding and reconstruction; For visual latent variables The motion prior feature map obtained by decoding and reconstruction.

7. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When one or more programs are executed by the one or more processors, the computing one or more processors implement a natural language motion prior-guided spatial faint target detection method as described in any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements a natural language motion prior-guided spatial faint target detection method as described in any one of claims 1 to 6.

Images