Method and system for identifying non-cooperative lunar and earth space targets based on improved deep learning model

Abstract

The application discloses a method and system for identifying non-cooperative targets in the earth-moon space based on an improved deep learning model, which is applied to the detection and matching of non-cooperative targets such as space debris and abandoned satellites in a complex background. The application first constructs a high-fidelity deep space non-cooperative target simulation dataset, simulates the observation data of various sensors in the earth-moon space environment, and accurately labels the data. Secondly, based on the YOLOv5 model, the self-attention mechanism and the spatial attention module are introduced. In view of the similarity problem of non-cooperative targets and background stars, the optical imaging data and laser ranging data are processed through a double-flow network architecture, and the cross-modal attention mechanism is introduced to optimize the feature extraction and target matching process and enhance the robustness of the model under low signal-to-noise ratio conditions. Moreover, the self-supervised contrast learning enhances the feature learning ability, can effectively distinguish non-cooperative targets from the background in a weak signal environment, and improves the real-time monitoring and early warning capability.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of space monitoring and deep learning technology, particularly applied in the field of space non-cooperative identification and matching. Specifically, the present application relates to an improved deep learning method, which can efficiently identify and match non-cooperative targets (such as space debris, abandoned satellites, etc.) in the space-earth space region, providing support for spacecraft orbit avoidance and mission planning BACKGROUND

[0002] With the increasing of human space activities, especially the increasing of space missions in low earth orbit and medium earth orbit regions, the problem of space debris has become a serious challenge. Space debris mainly refers to abandoned satellites, rocket debris and debris generated by collision events. These debris move at high speed in the earth's orbit, posing a great threat to existing spacecraft. According to data from NASA and other space agencies, the number of space debris in the earth's orbit has reached millions, and most of them are non-cooperative targets (NCT), i.e. targets that cannot be interacted with through traditional communication or control means. In the field of space, non-cooperative targets usually refer to space debris, abandoned satellites, unmarked small spacecraft, even extraterrestrial objects, etc. that do not operate according to predetermined rules. In the research of identification and orbit determination of non-cooperative targets in the space-earth space, a prominent technical difficulty is: how to effectively distinguish and match these non-cooperative targets without identification and active signal from numerous background stars under complex, sparse and low signal-to-noise ratio observation conditions.

[0003] Currently, the detection and identification of space debris mainly rely on radar, optical sensors and other observation equipment. However, these methods still have some shortcomings in practical application:

[0004] 1. Detection accuracy problem: Due to the small size, complex motion trajectory and high speed of space debris, traditional observation methods are difficult to provide accurate positioning and identification in a dynamically changing environment.

[0005] 2. Complex background interference: Space debris usually exists in a high background noise environment, such as strong sunlight, or affected by other satellites and celestial bodies, which is easily disturbed, resulting in a high misidentification rate of traditional methods.

[0006] 3. Real-time problem: The rapid movement of space debris requires the monitoring system to have real-time performance, while the existing technology mainly uses periodic scanning and offline processing method, which is difficult to meet the demand for efficient and real-time processing.

[0007] In recent years, with the development of deep learning technology, especially the successful application of deep learning models such as Convolutional Neural Networks (CNN) in image recognition, researchers have begun to try to use deep learning methods to improve the recognition accuracy and real-time performance of space debris. Deep learning can automatically learn and extract features, overcoming the limitations of manual feature design, and can handle recognition problems in complex backgrounds. However, existing deep learning-based space debris recognition methods still face the following challenges:

[0008] 1. Feature extraction problem: existing deep learning models have limited feature extraction capabilities for space debris, which cannot fully identify debris information in complex backgrounds, especially in low signal-to-noise ratio or high dynamic range.

[0009] 2. Efficient matching problem: the matching task of space debris not only requires high-precision feature extraction, but also requires efficient computing methods to compare debris and target databases. Existing technologies often face low computational efficiency and poor real-time performance.

[0010] Therefore, how to use deep learning technology to overcome these challenges, improve the recognition and matching accuracy of space debris, and achieve real-time monitoring and early warning, has become an important research direction in the field of space monitoring. SUMMARY

[0011] To overcome the shortcomings of the prior art, the present application provides a method and system for identifying non-cooperative targets in lunar and space based on an improved deep learning model, which improves the deep learning model YOLOv5 and introduces self-supervised learning to achieve efficient and accurate identification and matching of space debris in complex deep space environments, especially in cases where the background and target have high similarity.

[0012] According to one aspect of the present application, a method for identifying non-cooperative targets in lunar and space based on an improved deep learning model is provided, comprising:

[0013] Obtaining space debris images;

[0014] Inputting the obtained space debris images into a trained space non-cooperative target matching and recognition model to output the recognized non-cooperative targets; wherein the training of the space non-cooperative target matching and recognition model comprises:

[0015] Constructing a space debris dataset;

[0016] Based on the deep learning model YOLOv5, self-attention mechanism and spatial attention module are introduced, as well as multi-modal data fusion and cross-modal attention mechanism, forming an improved space non-cooperative target matching and recognition model;

[0017] Introduce self-supervised contrastive learning, and perform efficient feature extraction and lightweight model design on the improved space non-cooperative target matching recognition model to form an optimized space non-cooperative target matching recognition model.

[0018] According to the designed loss function, the optimized space non-cooperative target matching recognition model is trained on the constructed space debris data set to obtain a trained space non-cooperative target matching recognition model.

[0019] As a further technical solution, a space debris data set is constructed, including:

[0020] An orbit dynamics model is used to simulate the target trajectory in the space between the earth and the moon, considering the effects of multi-source gravitational field and non-spherical gravitational field, and providing high-precision position information and velocity distribution for the orbit evolution of the target under different initial conditions through numerical integration;

[0021] Highly realistic target optical imaging is generated in the simulation, including simulating the three-dimensional shape and surface reflection characteristics of the target, and batch data generation for different observation periods, observation angles and sensor parameters through a parallel rendering pipeline to form an original simulation data set;

[0022] The target positions in the original simulation data set are accurately labeled using known non-cooperative target trajectories and imaging geometry to form a space debris data set.

[0023] As a further technical solution, a self-attention mechanism and a spatial attention module are introduced, further including:

[0024] The self-attention mechanism calculates the correlation of each position in the feature map, dynamically adjusts the feature weight, and enhances the representation ability of the key area; the spatial attention module enhances the extraction of target features and strengthens the attention of the model to the target area, making the identification of small targets more accurate in complex environments.

[0025] As a further technical solution, a multi-modal data fusion is introduced, further including:

[0026] A dual-stream network structure is constructed to process optical imaging data and laser ranging data respectively, and data fusion is performed through a feature pyramid network and a path aggregation network to improve the integration capability of multi-modal data.

[0027] As a further technical solution, a cross-modal attention mechanism is introduced, further including:

[0028] The weights of different modal data are dynamically adjusted to enhance the adaptability of the model to different sensor data.

[0029] As a further technical solution, self-supervised contrastive learning is introduced to improve the efficient feature extraction and lightweight model design of the improved space non-cooperative target matching identification model, including:

[0030] Self-supervised contrastive learning is introduced to minimize the feature difference between different targets by enhancing the feature similarity of the same target under different perspectives.

[0031] Deep separable convolution is used to reduce the computational complexity and parameter quantity of the model.

[0032] Model pruning technology and lightweight network architecture design are adopted to optimize the YOLOv5 model so that it can run in real time in a resource-limited environment.

[0033] As a further technical solution, the loss function is designed, which also includes:

[0034] The focal loss is introduced to reduce the weight of easy-to-classify samples and focus on optimizing difficult-to-classify samples. The complete IoU loss is introduced to consider the overlap degree of the bounding box, the distance between the center points and the consistency of the aspect ratio, and to improve the accuracy of target boundary regression.

[0035] According to an aspect of the present application, a lunar space non-cooperative target recognition system based on an improved deep learning model is provided, comprising:

[0036] An image acquisition module is used to acquire space debris images.

[0037] An image recognition module is used to input the acquired space debris images into the trained space non-cooperative target matching identification model, and output the recognized non-cooperative target. The training of the space non-cooperative target matching identification model includes:

[0038] A space debris dataset is constructed.

[0039] Based on the deep learning model YOLOv5, self-attention mechanism and spatial attention module are introduced, and multi-modal data fusion and cross-modal attention mechanism are also introduced to form an improved space non-cooperative target matching identification model.

[0040] Self-supervised contrastive learning is introduced to improve the efficient feature extraction and lightweight model design of the improved space non-cooperative target matching identification model, forming an optimized space non-cooperative target matching identification model.

[0041] According to the designed loss function, the optimized space non-cooperative target matching identification model is trained on the constructed space debris dataset to obtain a trained space non-cooperative target matching identification model.

[0042] According to an aspect of the present application, there is provided a lunar space non-cooperative target recognition device based on an improved deep learning model, comprising a memory and a processor, the memory storing program instructions executed by the processor, and the processor calling the program instructions to execute the steps of the improved deep learning model-based lunar space non-cooperative target recognition method.

[0043] According to an aspect of the present application, there is provided a non-transitory computer readable storage medium storing computer instructions, which make the computer execute the steps of the improved deep learning model-based lunar space non-cooperative target recognition method.

[0044] Compared with the prior art, the present application trains a deep learning target detection and recognition model with strong robustness for non-cooperative targets and background stars in high-similarity scenarios based on high-similarity datasets. The main innovations and advantages of the present application are embodied in the following aspects:

[0045] 1. Deep learning combined with non-cooperative targets: The present application innovatively combines deep learning with non-cooperative targets and proposes a new non-cooperative target recognition method. By introducing a self-attention mechanism (Self-Attention) and a spatial attention module (Spatial Attention Module) into the model, the recognition ability of weak targets is enhanced, especially in low-contrast and complex background conditions, which can maintain high-precision target detection.

[0046] 2. Multi-modal data fusion and cross-modal attention mechanism: The present application proposes a multi-modal data fusion method, which designs a dual-stream network (Dual-Stream Network) structure to process optical imaging data and laser ranging data respectively, and uses a feature pyramid network (FPN) and a path aggregation network (PANet) for efficient fusion. Combined with the cross-modal attention mechanism (Cross-Modal Attention), the feature weights of different modalities are dynamically adjusted, thereby improving the adaptability of the model to different sensor data and ensuring efficient target recognition under various observation conditions. The advantage of this method is its powerful data processing capability, which can flexibly cope with complex and variable observation environments.

[0047] 3. Construction and application of high-fidelity simulation dataset: Based on the precise orbital dynamics model and the optical imaging law of non-cooperative targets, the invention constructs a high-fidelity simulation dataset. This dataset covers various target types, lighting conditions, noise interference, and complex backgrounds, greatly improving the training effect of deep learning models and providing rich data support for subsequent algorithm optimization and performance evaluation. This dataset is not only crucial for target recognition and trajectory prediction tasks, but also provides important reference for practical applications in deep space environment. BRIEF DESCRIPTION OF DRAWINGS

[0048] To more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the drawings used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0049] Figure 1 The flowchart of the method for identifying non-cooperative targets in lunar and space based on the improved deep learning model provided by the embodiments of the present application.

[0050] Figure 2 The training flowchart of the space non-cooperative target matching identification model provided by the embodiments of the present application.

[0051] Figure 3 The flowchart of constructing space debris dataset provided by the embodiments of the present application.

[0052] Figure 4 The space non-cooperative target dataset part sample diagram provided by the embodiments of the present application.

[0053] Figure 5 The high-simulation space non-cooperative target test set evaluation part result visualization diagram provided by the embodiments of the present application. DETAILED DESCRIPTION

[0054] The present application proposes a method for identifying non-cooperative targets in lunar and space based on an improved deep learning model, which is particularly suitable for detecting and matching non-cooperative targets (such as space debris, retired spacecraft, etc.) in complex backgrounds under long exposure time and star tracking mode. This method takes YOLOv5 as the basic model, combines the characteristics of deep space environment and multiple innovative technologies, and significantly improves the accuracy and efficiency of non-cooperative target recognition and matching in complex backgrounds, mainly in:

[0055] Firstly, the present application firstly constructs a high-fidelity simulation dataset to accurately simulate the non-cooperative target in the lunar space environment. By using the orbital dynamics model and the space environment characteristics of the lunar space, the target observation process of multiple types of sensors (such as ground-based optical telescopes, on-orbit cameras, laser range finders, etc.) under different observation conditions is simulated. In the simulation, factors such as the optical imaging law of the target, the distribution of the background star field, the motion blur effect, and the density of the star field are considered to ensure that the appearance characteristics and background similarity of the target reach a high degree of fidelity to the real scene. In addition, through an automatic labeling mechanism, the accurate trajectory, appearance attribute, and orbit parameters of the target are generated for each piece of data, ensuring the high quality of the training data.

[0056] Secondly, the present application develops a single-stage deep learning detection framework. In view of the high similarity between the non-cooperative target and the background star in visual features, the YOLOv5 model is adopted, and deep customization and optimization are carried out based on it. The specific innovations include: the introduction of self-attention mechanism (Self-Attention Mechanism) and spatial attention module (Spatial Attention Module), which enables the model to capture weak target features more finely and improve the sensitivity of target detection. Through a double-flow network structure, optical imaging data and laser ranging data are processed respectively, and a feature fusion layer is set between the feature pyramid network (FPN) and the path aggregation network (PANet), which improves the integration ability of different modal data. Cross-modal attention mechanism (Cross-Modal Attention) is introduced, which can dynamically adjust the weight of different modal data, enhance the adaptability of the model to multiple data sources, and improve the detection accuracy of non-cooperative targets. In addition, self-supervised methods such as contrastive learning (Contrastive Learning) are used to enhance the robustness of the model under weak features and low signal-to-noise ratio conditions, so that the model can distinguish targets from backgrounds in visually indistinguishable environments by relying on deep features.

[0057] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of protection of the present application. In addition, the technical features in each embodiment or in a single embodiment provided by the present application can be combined with each other at will to form new technical solutions, and such combination is not restricted by the order of steps and / or structure composition mode, but should be based on the realization by those skilled in the art. When the combination of technical solutions appears contradictory or cannot be realized, it should be considered that the combination of technical solutions does not exist and is not within the scope of protection required by the present application.

[0058] Referring to Figure 1 The embodiment of the present application provides a lunar-space non-cooperative target recognition method based on an improved deep learning model. First, a space debris image is obtained. Then, the obtained space debris image is input into a trained space non-cooperative target matching recognition model to output the recognized non-cooperative target.

[0059] Further, the space debris picture is input into the space non-cooperative target matching recognition model, the optical imaging data and the laser ranging data in the space debris picture are respectively subjected to feature extraction, and a self-attention mechanism and a spatial attention module are introduced during feature extraction. Then, the spatial attention images on the two branches are subjected to multi-modal data fusion, and a cross-modal attention mechanism is introduced during multi-modal data fusion. Finally, the recognized non-cooperative target is output.

[0060] Referring to Figure 2 The training and evaluation of the space non-cooperative target matching recognition model in the embodiment of the present application include:

[0061] Step 1: Construct and preprocess the space debris dataset, including obtaining space debris image data of different types, and performing image synthesis and data labeling;

[0062] Step 2: Improve the YOLOv5 model, introduce a self-attention mechanism (Self-Attention Mechanism) and a spatial attention module (Spatial Attention Module) to the YOLOv5 model, and innovatively introduce a multi-modal data fusion technology and a cross-modal attention mechanism (Cross-Modal Attention);

[0063] Step 3: The loss function of the improved YOLOv5 is customized and designed to better cope with the class imbalance and target boundary blur problems in deep space environment;

[0064] Step four: A self-supervised contrastive learning method was introduced to the optimized YOLOv5 model, and efficient feature extraction and lightweight model design were performed.

[0065] Step five: The optimized YOLOv5 model was trained and evaluated for accuracy.

[0066] Please refer to Figure 3 , the construction and preprocessing of the space debris dataset in step one further includes:

[0067] 1. Dataset construction and target simulation

[0068] Orbit simulation: First, according to the distribution and motion characteristics of non-cooperative targets (such as space debris, retired spacecraft, and small unknown targets) in the Earth-Moon space, a dynamics model is established that includes multi-source gravitational fields of the Earth, the Moon, the Sun, and non-spherical gravitational field effects. This model provides orbit evolution information for targets under different initial conditions through numerical integration and parameter iteration.

[0069] Imaging simulation: The imaging process of actual observation sensors (such as ground-based optical telescopes and on-orbit cameras) is simulated through a high-precision optical imaging simulation framework, considering factors such as atmospheric disturbance, optical system aperture, detector noise, thermal noise, and photon counting noise.

[0070] Background star field generation: High-fidelity starry night backgrounds are generated using star catalog data to simulate star scenes of different magnitudes and densities, ensuring that the background and target have high similarity in visual features.

[0071] Target material and shape modeling: Through three-dimensional CAD modeling and optical BRDF (Bidirectional Reflectance Distribution Function) models, the material, shape, and reflection characteristics of the target are considered to ensure that the target appears similar to a star in the image.

[0072] In the above steps, first, for the distribution and motion characteristics of non-cooperative targets in the Earth-Moon space environment, a dynamic model is established that includes the effects of multi-source gravitational fields such as the Earth, the Moon, the Sun, and non-spherical gravitational fields. This model provides high-precision position information and velocity distribution for the target's orbit evolution under different initial conditions through numerical integration and parameter iteration. The focus of this step is to ensure the diversity and authenticity of the target trajectory: we set different orbit inclination, semi-major axis, eccentricity, and relative illumination conditions according to various possible target types (space debris, retired spacecraft, micro unknown targets), so that the target in the simulation is as close as possible to the actual deep space environment. After obtaining the trajectory information of the non-cooperative target, we simulate the imaging process of the actual observation sensor (such as ground-based medium aperture optical telescope, on-orbit camera) under specific attitude, detection sensitivity, and field of view conditions using an optical imaging simulation framework. Here, high-precision optical transmission models and atmospheric disturbance estimation models are introduced, considering the comprehensive effects of optical system aperture size, detector pixel size, readout noise, photon counting noise, thermal noise, and observation conditions (such as sky background brightness, moonlight interference, detection wavelength, etc.). At the same time, we use star catalog data to generate a vast and realistic starry sky background, simulating different star magnitude levels and different density distributions of star scenes, so that the appearance characteristics of background light sources and potential targets in brightness, star image size, and gray scale distribution are highly similar to real observation data. This not only increases the difficulty of data generation and processing, but also significantly improves the simulation degree and scientific value of the data set. To highlight the proximity of non-cooperative targets and background stars in visual features, we consider the target material, shape structure, and surface reflection characteristics in the target simulation. By constructing a three-dimensional CAD model of the target and fitting the corresponding optical BRDF (Bidirectional Reflectance Distribution Function) model, we perform lighting analysis and image rendering for targets of different shapes, sizes, and materials. Due to the lack of specific identification, the visible image points of the target are similar to point sources or near-point source star images, with gray scale distribution characteristics and apparent size similar to surrounding stars. This detailed modeling process significantly increases the difficulty and workload of data set construction, but also makes the resulting data more challenging and representative in subsequent deep learning training.

[0073] 2. Data augmentation and annotation

[0074] Batch generation of images with different observation periods, angles, and sensor parameters using parallel rendering pipelines to simulate various possible disturbances and noises. Accurate annotation of targets in images, including target position, bounding box coordinates, relative brightness information, etc. Ensure the accuracy of the annotation information to provide high-quality data for subsequent model training.

[0075] Specifically, after the trajectory calculation and optical simulation are completed, we use the parallel rendering pipeline to generate batch data for different observation periods, observation angles, and observation sensor parameters. During this process, various noise models, lighting changes, scattering effects, and possible motion blurs are also introduced. To ensure data diversity, we have carried out large-scale randomization in parameter selection, including but not limited to target distance distribution, relative azimuth angle variation, random disturbance of background star field density and luminosity, and moderate changes in observation period (dawn, dusk, night).

[0076] After the above multi-step, high-precision, high-computational simulation and data synthesis, we finally generate a large-scale and highly simulated deep space non-cooperative target dataset, as shown in Figure 4 The dataset consists of 6000 simulated optical observation images in deep space environment. In each image, the presentation position, brightness characteristics, and star size of the non-cooperative target are carefully designed to be highly similar to the background stars in visual features. The basic elements of each image in the dataset include: 1) background star field: containing stars of different magnitudes, with brightness and density ranges generated according to real star catalog data. 2) non-cooperative target: one or several targets scattered in the image, with similar appearance characteristics to star point sources, but can be distinguished by motion characteristics in continuous multi-time images. 3) observation noise and imaging distortion: simulation of sensor noise, optical system aberration, field edge dimming, and star point spread function (PSF) variation to provide realistic actual disturbance scenarios for subsequent algorithm development.

[0077] After the construction of the original simulation data, we use the known non-cooperative target trajectory and imaging geometry to accurately label the target positions in the dataset. Through automated labeling tools and post-inspection, we generate standardized labeling information for each non-cooperative target in the image (such as target position bounding box coordinates, circular ROI, target relative brightness information, and orbit parameter labels). The labeling process is rigorous and tedious, ensuring that the labeling results are free of missing labels, incorrect labels, and ambiguity, providing high-quality reference standards for supervised training of deep learning models. After completing the labeling, we divide the dataset into a reasonable division suitable for deep learning training process. Randomly select 1000 images from the 6000 images as the test set, and the remaining 5000 images as the training and validation set. During training, a portion of the training set can be further divided as a validation set for parameter tuning and overfitting detection in model training. This data division method ensures that the model can fully learn the subtle distinguishing features between non-cooperative targets and background stars in large-scale, complex and diverse training data, while the independent test set objectively tests the generalization performance of the final model.

[0078] 3. Dataset division and preparation

[0079] The 6000 image dataset is divided into training set, validation set and test set, of which 1000 is independent test set, and the remaining 5000 is used for training and validation. With automatic labeling tool and manual inspection, ensure that the labeling result is correct, no missing or ambiguity, and divide the data for training and validation of deep learning model.

[0080] The improved YOLOv5 model design in step two further includes:

[0081] The self-attention mechanism and spatial attention module are introduced. The self-attention mechanism enhances the model's attention to target area features, improving the detection ability of weak targets. This mechanism calculates the correlation of each position in the feature map, dynamically adjusts the feature weight, and enhances the representation ability of key areas. The spatial attention module further enhances the extraction of target features, by strengthening the model's attention to target areas, making the identification of small targets more accurate in complex environments.

[0082] Multi-modal data fusion and cross-modal attention mechanism are introduced, and a double-flow network structure is designed to process optical imaging data and laser ranging data respectively. Through the feature fusion layer in the feature pyramid network (FPN) and the path aggregation network (PANet), efficient integration of different modal features is achieved. Cross-modal attention mechanism is introduced to dynamically adjust the weight of different modal features, further improving the accuracy and robustness of target detection.

[0083] The second step model design and training selects a YOLOv5 model architecture suitable for space debris recognition. First, we deeply customized the YOLOv5 model to adapt to the high similarity of non-cooperative targets and background stars in visual features in deep space environment. Traditional YOLOv5 model mainly faces target detection in ground or low orbit environment, and its performance is lacking when dealing with high noise, low contrast and high similarity between target and background. Therefore, we introduced self-attention mechanism and spatial attention module in the feature extraction part of the model to enhance the model's ability to capture weak target features and improve the precision of feature expression. Secondly, we innovatively introduced multi-modal data fusion technology to effectively combine optical imaging data and laser ranging data. By designing a double-flow network structure, different modal data are processed, and a feature fusion layer is set between the feature pyramid network (FPN) and the path aggregation network (PANet) to realize the efficient integration of different modal features. Further, the application of cross-modal attention mechanism enables the model to dynamically adjust the weight of each modal feature, improving the accuracy and robustness of target detection.

[0084] Data augmentation and feature extraction: In deep space environment, non-cooperative targets and background stars are highly similar in visual features such as brightness, shape, and gray scale distribution. Single modal optical data faces significant detection challenges. In order to effectively distinguish these highly similar targets and backgrounds, we have designed and implemented a number of innovative optical data enhancement and feature extraction methods on the basis of YOLOv5 model. These innovative methods significantly improve the detection accuracy and robustness of the model in complex deep space environment, solving the limitations of traditional methods in dealing with low signal-to-noise ratio and high similarity background.

[0085] To enhance the model's ability to capture weak target features, we introduced the Self-Attention Mechanism and Spatial Attention Module in the feature extraction part of YOLOv5. Self-Attention Mechanism is a widely used technique in deep learning, especially when dealing with sequential data. It allows the model to dynamically consider information from other parts of the input when processing a certain part, thus improving the model's expressive power. Self-Attention Mechanism is one of the core components of the Transformer model and is widely used in Natural Language Processing (NLP) and Computer Vision. The core idea of Self-Attention Mechanism is to allow each element of the input data to adjust its representation based on the relationship with other elements. Specifically, each element in the input sequence is compared with other elements, and information is combined based on this comparison. This allows the model to capture dependencies between different positions in the input sequence.

[0086] Self-Attention Mechanism is typically implemented through the following steps: the input is a sequence of length n, assuming the dimension of each element is d, the input sequence can be represented as an n x d matrix. Each input element will generate a query (Query), key (Key) and value (Value) through three different linear transformations (matrix multiplication). These transformations are learned through training. For the i-th input element x i , its query, key and value are:

[0087] (1)

[0088] where Q, K, V are the Query, Key, Value matrices, and W K and W V are weight matrices that need to be learned. The dot product of the query and the key is used to calculate the attention score, which reflects the correlation between input elements:

[0089] (2)

[0090] where is the dimension of the key vector, and the score is usually scaled to avoid large numerical values. Self-Attention Mechanism enhances the model's focus on key regions through global feature interaction, improving the detection ability of weak targets. Self-Attention Mechanism is a powerful tool that can help the model learn the dependencies between different parts of the input data and is widely used in Natural Language Processing and Computer Vision. Its advantages include capturing global dependencies, strong parallelization capabilities, and high processing flexibility, making it an important innovation in deep learning.

[0091] Spatial Attention Module (SAM) is a variant of self-attention mechanism, mainly used in image processing tasks. It enhances the information of specific regions by focusing on the weight distribution of spatial positions. The goal of SAM is to focus on the most important regions in the image and ignore irrelevant parts by applying attention weighting in the spatial dimension, thereby improving the performance of the model.

[0092] In many computer vision tasks, important regions in the image are usually not evenly distributed. SAM can help the model automatically find these key regions, similar to how the human eye tends to focus on the most visually prominent parts when observing an image, thereby enhancing the expressive ability of image features. SAM is often used in conjunction with Channel Attention Module (CAM) to form a more refined attention mechanism. The basic principle of SAM is as follows:

[0093] Input Feature Map: Given a feature map F ∈ R H×W×C where H is the height of the image, W is the width of the image, and C is the number of channels. The purpose of SAM is to weight the spatial dimension (i.e., H × W) of the feature map.

[0094] Generate Spatial Attention Map: The core idea of SAM is to learn a spatial attention map based on the features of different positions in the image. This process usually includes the following steps:

[0095] Compress the input feature map along the channel dimension through some operations (such as max pooling, average pooling, etc.) to generate a two-dimensional spatial feature map (i.e., a H × W feature map). The purpose of this step is to extract the global features of each spatial position.

[0096] Merge the generated spatial feature map (e.g., concatenate or weight) and then pass it through a convolution layer (usually a 1 × 1 convolution) to generate a spatial attention map. The size of the spatial attention map is H × W, and the value at each position represents the importance of that position.

[0097] Specifically, assume that the feature maps obtained by max pooling and average pooling are Fmax and Favg, respectively. After merging, the spatial attention map S is calculated through convolution:

[0098]

[0099] where, is an activation function (such as sigmoid), Conv1 is a 1x1 convolution operation, and Concat represents a concatenation operation. After obtaining the spatial attention map, it is multiplied element-wise with the original input feature map. In this way, the model can weight the input feature map according to the weight of each spatial position, thereby strengthening or suppressing the information in a specific region.

[0100]

[0101] where Fout is the weighted output feature map and S is the spatial attention map. The spatial attention module helps the model focus on important regions in the image by learning and weighting the spatial information in the image, thereby enhancing feature representation and improving the performance of computer vision tasks. When combined with the channel attention module, the spatial attention module can more comprehensively enhance the attention mechanism of the model, making it more flexible and efficient in processing images.

[0102] To enable the model to dynamically adjust the weights of each modality feature and improve the accuracy and robustness of target detection, a cross-modal attention mechanism is introduced. Cross-modal attention mechanism is a mechanism for handling relationships between multi-modal data such as images, text, audio, etc. Its goal is to better integrate multi-modal information by establishing interdependence and relationships between different modalities, thereby improving the performance of multi-modal tasks such as visual question answering, image description, speech recognition, etc.

[0103] The core idea of cross-modal attention mechanism is to pass information and fuse between different modalities of data through attention mechanism. Specifically, suppose we have two different modalities of data (for example, image and text), we want to capture the associated information between image and text through attention mechanism and fuse them into a joint representation.

[0104] Input feature representation: Suppose we have two modalities of data: one is image feature V∈R Hv×Wv×Cv (may be a feature map extracted by a convolutional neural network), and the other is text feature T∈R Nt×Dt (may be a word embedding or a sentence feature extracted by an RNN, Transformer, etc.). Here Hv, Wv is the height and width of the image, Cv is the number of channels of the image feature map, Nt is the length of the text, and Dt is the dimension of the text feature.

[0105] Generating Query, Key, and Value: To enable cross-modal attention between different modalities, we first take the features of one modality (e.g., text features) as the query (Query), and the features of another modality (image features) as the key (Key) and value (Value). For image features V, we usually generate the key and value through some linear transformation. Assuming that each pixel point (or region) of the image feature map represents information at a location, we want to find related information from the text. For text features T, we generate a query through linear transformation to obtain related information from image features.

[0106] Computing Cross-Modal Attention Weights: Assuming we take text features as the query and image features as the key and value. First, calculate the similarity between the text query and the image key to get the relationship between each text element and image element:

[0107]

[0108] where Q is the text feature, K is the image feature, d k is the dimension of the key vector. Then, convert this similarity into attention weights through the softma function:

[0109]

[0110] This weight reflects the correlation between the i-th word in the text and the j-th region in the image. Weighted summation of information: After obtaining the attention weights, the value of the image is weighted and summed through these weights to obtain a new image feature representation as a supplement to the text information:

[0111]

[0112] Similarly, we can also calculate the relationship between image features and text features by taking the image as the query and the text as the key and value.

[0113] Multi-modal fusion: The weighted features obtained through the cross-modal attention mechanism will contain information from both modalities. Typically, these features will be further fused (e.g., concatenation, summation, etc.) and used for subsequent tasks (such as classification, generation, etc.).

[0114] Cross-modal attention mechanism is a very powerful tool that can help different modalities of data establish effective association and information transmission. It is widely used in visual question answering, image description generation, multi-modal sentiment analysis, and other tasks, significantly improving the performance of multi-modal learning. By establishing interdependent relationships between different modalities, cross-modal attention mechanisms not only enhance the accuracy of information fusion, but also improve the model's ability to handle complex tasks.

[0115] The loss function of the improved YOLOv5 in step three is customized and designed, further comprising:

[0116] The customized loss function design of the model is to cope with the class imbalance and target boundary ambiguity in deep space environment. The focal loss is introduced to reduce the weight of easy classification samples and optimize difficult classification samples, thereby improving the detection accuracy. The complete IoU loss (CIoU Loss) is introduced to consider the overlap degree of the boundary box, the distance between the center points and the consistency of the aspect ratio, which effectively improves the accuracy of target boundary regression.

[0117] The loss function of the improved YOLOv5 in step three is customized and designed to better cope with the class imbalance and target boundary ambiguity in deep space environment. The customized loss function design: for the class imbalance and target boundary ambiguity in deep space non-cooperative target detection, we customize the loss function of YOLOv5, introduce focal loss (Focal Loss) and complete IoU loss (Complete IoU Loss, CIoU Loss) to better adapt to the special needs of deep space environment.

[0118] The expression of focal loss (Focal Loss) is:

[0119]

[0120] wherein, is the predicted probability of the model for the true class, and is a hyperparameter. Focal loss reduces the loss weight of easy classification samples and optimizes difficult classification samples, significantly reducing the false detection rate.

[0121] The expression of complete IoU loss (Complete IoU Loss, CIoU Loss) is:

[0122]

[0123] wherein, represents the Euclidean distance between the center points, c is the diagonal length of the smallest closed frame enclosing the two frames, v is the aspect ratio consistency measure, is a weight coefficient. CIoU loss not only considers the overlap degree of the boundary box, but also includes the distance between the center points and the consistency of the aspect ratio, so that the boundary box regression is more accurate, further improving the detection accuracy.

[0124] The expression of the comprehensive loss function is:

[0125]

[0126] wherein, are the weight coefficients of each loss term. By customizing the loss function, the learning ability of the model in the complex deep space environment is optimized, and the detection performance is significantly improved.

[0127] In step four, the optimized YOLOv5 model is introduced into the self-supervised contrastive learning method, and efficient feature extraction and lightweight model design are performed, further including:

[0128] The self-supervised contrastive learning method is introduced, which enhances the feature similarity of the same target under different perspectives and minimizes the feature difference between different targets, enhancing the model's discrimination ability in complex environments. Depthwise separable convolution is used to reduce the computational complexity and parameter quantity of the model, improving the feature extraction efficiency. Model pruning technology and lightweight network architecture design are used to optimize the YOLOv5 model, enabling it to run in real-time in resource-constrained environments such as deep space probes or satellites.

[0129] Step four introduces the self-supervised contrastive learning method and performs efficient feature extraction and lightweight model design. Self-supervised contrastive learning: In view of the sparsity and diversity of deep space observation data, a series of data enhancement techniques are adopted, and the self-supervised contrastive learning method is introduced to improve the feature expression ability and generalization performance of the model.

[0130]

[0131] wherein, represents the similarity of the features of two samples, τ is the temperature parameter, and N is the batch size. By maximizing the feature similarity of the same target under different perspectives and minimizing the feature difference between different targets, the model's discrimination ability in complex deep space environments is enhanced.

[0132] Efficient feature extraction and lightweight model design: To meet the high requirements of deep space target detection on real-time performance and computing resources, efficient feature extraction and lightweight design are performed on the YOLOv5 model to optimize the computational efficiency and inference speed of the model. Depthwise separable convolution:

[0133]

[0134] By decomposing the standard convolution into depthwise convolution and pointwise convolution, the model parameter quantity and computational complexity are significantly reduced, and the feature extraction efficiency is improved. Model pruning: through pruning technology, redundant convolution kernels and neurons in the model are removed, further reducing the computational resource demand of the model and improving the inference speed. Lightweight network architecture design: adopt a bottleneck structure similar to MobileNet, optimize the structural layout of the Backbone network, improve the efficiency and effect of feature extraction, and ensure that the model can run in real time on resource-limited deep space probes or satellites.

[0135] The model training and model evaluation in step five further include:

[0136] 1. Model training

[0137] The improved YOLOv5 model is trained using a high-fidelity simulation dataset. The training process uses a batch size of 16, an initial learning rate of 1e-3, and a cosine annealing learning rate scheduling strategy. To prevent overfitting, the Early Stopping strategy is used, and when the F1 score on the validation set does not improve for 10 consecutive epochs, the training is terminated early.

[0138] 2. Model evaluation and precision verification

[0139] After training is complete, the model is evaluated using an independent test set. As shown in Figure 5 the test results show that the improved YOLOv5 model has excellent space debris recognition accuracy and recall rate in deep space environments: the recall rate reaches 96%, and the precision rate reaches 93%.

[0140] The step five trains and evaluates the precision of the optimized YOLOv5 model. In this invention, based on the YOLOv5 model and the constructed high-fidelity simulation deep space non-cooperative target dataset, we carried out the recognition and matching experiment of deep space non-cooperative targets to verify the effectiveness and robustness of the developed deep learning algorithm. During the experiment, 1000 of the 6000 simulation images were randomly selected as an independent test set, and the remaining 5000 were used for model training and verification. The model training used a batch size of 16, an initial learning rate of 1e -3and combined with cosine annealing learning rate scheduling strategy to ensure the model can effectively converge during training. To prevent overfitting, we introduced the Early Stopping strategy, which terminates training early when the F1 score on the validation set does not improve for 10 consecutive epochs. During training, the customized loss function design, including Focal Loss and Complete IoU Loss (CIoU Loss), played a key role in addressing class imbalance and target boundary ambiguity issues, significantly improving the model's detection accuracy.

[0141] Table 1 Evaluation results of high-fidelity spatial non-cooperative target test set

[0142]

[0143] After completing the model training, we evaluated the model on an independent test set, and the results are shown in Table 1. The model's recall rate (recognition rate) reached 96%, and the precision rate (matching rate) was 93%. This high recall rate indicates that the model can effectively identify most real deep space non-cooperative targets, greatly reducing the likelihood of missed detection. This is crucial for monitoring and managing the Earth-Moon space environment, ensuring that most space debris and abandoned spacecraft are discovered in a timely manner, reducing the risk of on-orbit spacecraft collisions. Meanwhile, a precision rate of 93% means that 93% of the identified targets are true non-cooperative targets, with a false detection rate of only 7%. This high precision effectively reduces resource waste and subsequent processing burden due to false detection, improving the overall efficiency and reliability of the system.

[0144] Based on the aforementioned training process, the invention can efficiently and accurately identify and match space debris in complex deep space environments, especially in situations where the background and target similarity is high. The invention not only improves the detection accuracy of space debris but also reduces the false detection rate, providing strong technical support for deep space exploration and space debris management.

[0145] The application is based on a deep space non-cooperative target recognition and matching algorithm developed by YOLOv5. Through the introduction and implementation of multiple innovative technologies, the problem of high similarity between target and background stars in visual features in deep space environment is successfully solved, and the detection accuracy, robustness and real-time performance of the model are significantly improved. This innovative algorithm system not only provides an efficient and reliable technical means for automatic detection and matching of non-cooperative targets in the Earth-Moon space, but also has important scientific significance and wide application prospects in promoting the application and development of deep learning technology in the field of space science. In the future, we will further optimize the algorithm structure, improve its adaptability and real-time performance in actual deep space observation data, and explore more fusion methods of deep learning and orbit dynamics model, continuously improve the recognition and orbit determination capability of non-cooperative targets, and provide more reliable technical support for the safety and sustainable development of space environment. Through the above steps, the application provides a space debris matching and recognition method based on the YOLOv5 model, which can effectively improve the automatic recognition and orbit prediction accuracy of space debris, has high real-time performance and accuracy, and is suitable for deep space exploration, satellite collision avoidance and other fields.

[0146] The implementation basis of each embodiment of the application is achieved by programmed processing of a device with processor function. Therefore, in engineering practice, the technical solutions and functions of each embodiment of the application are packaged into various modules. Based on this actual situation, on the basis of the above embodiments, the embodiment of the application provides a non-cooperative target recognition system in the Earth-Moon space based on an improved deep learning model, which is used to execute the non-cooperative target recognition method in the Earth-Moon space based on the improved deep learning model in the above method embodiment.

[0147] The system comprises an image acquisition module for acquiring space debris images, and an image recognition module for inputting the acquired space debris images into a trained space non-cooperative target matching and recognition model and outputting recognized non-cooperative targets. The training of the space non-cooperative target matching and recognition model comprises: constructing a space debris dataset; introducing a self-attention mechanism and a spatial attention module on the basis of a deep learning model YOLOv5, and introducing multi-modal data fusion and cross-modal attention mechanism to form an improved space non-cooperative target matching and recognition model; introducing self-supervised contrast learning to perform efficient feature extraction and lightweight model design on the improved space non-cooperative target matching and recognition model to form an optimized space non-cooperative target matching and recognition model; and training the optimized space non-cooperative target matching and recognition model on the constructed space debris dataset according to a designed loss function to obtain a trained space non-cooperative target matching and recognition model.

[0148] The system for identifying non-cooperative targets in the earth-moon space based on the improved deep learning model provided by the embodiment of the application faces the demand for improving the identification and matching accuracy of space debris in the current space monitoring field and realizing real-time monitoring and early warning, adopts the foregoing modules, takes YOLOv5 as a basic model, solves the detection and matching problem of non-cooperative targets (such as space debris, retired aircraft, etc.) in a complex background under a long exposure time and a star tracking mode through model improvement and high-fidelity dataset construction, and improves the accuracy and efficiency of identification and matching of non-cooperative targets in a complex background.

[0149] It should be noted that the system embodiments provided by the application are used to implement the methods in the method embodiments and the methods in other method embodiments provided by the application, and the difference is only that the corresponding function modules are set, the principle is basically the same as that of the above-mentioned system embodiments provided by the application, as long as the person skilled in the art improves the modules in the above-mentioned system embodiments on the basis of the above-mentioned system embodiments, refers to the specific technical solutions in other method embodiments, obtains the corresponding technical means through the combination of technical features, and the technical solutions composed of these technical means, as long as the technical solutions have practicality, the corresponding system embodiments are obtained by improving the modules in the above-mentioned system embodiments, and are used to implement the methods in other method embodiments.

[0150] Based on the same inventive concept as the foregoing embodiments, the embodiment of the application further provides a device for identifying non-cooperative targets in the earth-moon space based on an improved deep learning model, comprising a memory and a processor, the memory stores program instructions executed by the processor, and the processor calls the program instructions to execute the steps of the method for identifying non-cooperative targets in the earth-moon space based on the improved deep learning model.

[0151] Based on the same inventive concept as the foregoing embodiments, the embodiment of the application further provides a non-transitory computer readable storage medium, which stores computer instructions, and the computer instructions make the computer execute the steps of the method for identifying non-cooperative targets in the earth-moon space based on the improved deep learning model.

[0152] In summary, the application discloses a method for identifying non-cooperative targets in the earth-moon space based on an improved deep learning model, which is particularly applied to the detection and matching of non-cooperative targets such as space debris and abandoned satellites in a complex background. In view of the current situation that the detection and identification of non-cooperative targets in low earth orbit and earth-moon space face challenges such as strong background interference and low signal-to-noise ratio, the application proposes an efficient target matching method based on the YOLOv5 model. By introducing a self-attention mechanism, a spatial attention module and a multi-modal data fusion technology, the recognition and matching accuracy of non-cooperative targets in a complex background is significantly improved. Specifically, the application first constructs a high-fidelity deep space non-cooperative target simulation dataset by simulating the observation data of various sensors in the earth-moon space environment and accurately labeling the data. Secondly, based on the YOLOv5 model, a self-attention mechanism and a spatial attention module are introduced. In view of the similarity problem between non-cooperative targets and background stars, the application processes optical imaging data and laser ranging data through a double-flow network architecture and introduces a cross-modal attention mechanism to optimize the feature extraction and target matching process and enhance the robustness of the model under low signal-to-noise ratio conditions. Through self-supervised contrastive learning, the application can effectively distinguish non-cooperative targets from the background in a weak signal environment and improve the real-time monitoring and early warning capability. The technical solution of the application can provide efficient and accurate support for spacecraft orbit avoidance and mission planning, and has a wide application prospect and good market prospect.

[0153] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the application, and not to limit them; although the application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the technical solutions of the application.

Claims

1. A method for identifying a non-cooperative target in a lunar-moon space based on an improved deep learning model, characterized in that, The method comprises the following steps: acquiring a space debris image; inputting the acquired space debris image into a trained space non-cooperative target matching and recognition model to output a recognized non-cooperative target; wherein the training of the space non-cooperative target matching and recognition model comprises the following steps: constructing a space debris dataset; on the basis of a deep learning model YOLOv5, introducing a self-attention mechanism and a space attention module, further introducing multi-modal data fusion and a cross-modal attention mechanism to form an improved space non-cooperative target matching and recognition model; the introduction of the self-attention mechanism and the space attention module further comprises: calculating the correlation of each position in the feature map through the self-attention mechanism, dynamically adjusting the feature weight, and enhancing the representation ability of the key region; through the space attention module, the extraction of target features is enhanced, and the attention of the model to the target region is strengthened, so that the recognition of the micro target in the complex environment is more accurate; introducing self-supervised contrast learning to perform efficient feature extraction and lightweight model design on the improved space non-cooperative target matching and recognition model to form an optimized space non-cooperative target matching and recognition model; training the optimized space non-cooperative target matching and recognition model on the constructed space debris dataset according to the designed loss function to obtain a trained space non-cooperative target matching and recognition model.

2. The method according to claim 1, wherein, The space debris dataset is constructed, comprising: using an orbital dynamics model to simulate the trajectory of a target in the earth-moon space, considering the effects of multi-source gravitational field and non-spherical gravitational field, and providing high-precision position information and velocity distribution for the orbital evolution of the target under different initial conditions through numerical integration; generating highly realistic target optical imaging in simulation, including simulating the three-dimensional morphology and surface reflection characteristics of the target, and generating batch data for different observation periods, observation angles and sensor parameters through a parallel rendering pipeline to form an original simulation dataset; using the known trajectory of the non-cooperative target and the imaging geometry, accurately labeling the target position in the original simulation dataset to form a space debris dataset.

3. The method of claim 1, wherein the improved deep learning model is based on a deep convolutional neural network. introducing multi-modal data fusion further comprises: constructing a double-flow network structure to process optical imaging data and laser ranging data respectively, and performing data fusion through a feature pyramid network and a path aggregation network to improve the integration ability of multi-modal data.

4. The method according to claim 3, wherein, introducing a cross-modal attention mechanism further comprises: dynamically adjusting the weight of different modal data to enhance the adaptability of the model under different sensor data.

5. The method of claim 1, wherein the improved deep learning model is based on a method of identifying a non-cooperative target in a lunar-moon space. introducing self-supervised contrast learning for efficient feature extraction and lightweight model design of the improved space non-cooperative target matching and recognition model comprises: introducing self-supervised contrast learning to minimize the feature difference between different targets by enhancing the feature similarity of the same target under different perspectives; using a depth separable convolution to reduce the computational complexity and parameter quantity of the model; adopting model pruning technology and lightweight network architecture design to optimize the YOLOv5 model so that it can run in real time in a resource-limited environment.

6. The method of claim 1, wherein the improved deep learning model is based on a method of identifying a non-cooperative target in a lunar-moon space. designing a loss function further comprises: The focal loss is introduced to reduce the weight of easy-to-classify samples and focus on optimizing difficult-to-classify samples; the complete IoU loss is introduced to consider the overlap degree of the bounding box, the distance between the center points and the consistency of the aspect ratio, and improve the accuracy of target boundary regression.

7. A lunar and earth space non-cooperative target recognition system based on an improved deep learning model, characterized in that, The method comprises the following steps: An image acquisition module is configured to acquire space debris images. An image recognition module is configured to input the acquired space debris images into a trained space non-cooperative target matching and recognition model, and output recognized non-cooperative targets. The training of the space non-cooperative target matching and recognition model comprises the following steps: A space debris dataset is constructed. On the basis of a deep learning model YOLOv5, a self-attention mechanism and a space attention module are introduced, and a multi-modal data fusion and a cross-modal attention mechanism are also introduced to form an improved space non-cooperative target matching and recognition model. The self-attention mechanism and the space attention module also comprise the following steps: the correlation of each position in a feature map is calculated through the self-attention mechanism, the feature weight is dynamically adjusted, and the representation ability of a key region is enhanced; the extraction of target features is enhanced through the space attention module, the attention of the model to the target region is strengthened, and the identification of a small target in a complex environment is more accurate.

8. A device for identifying a non-cooperative lunar-solar space target based on an improved deep learning model, characterized in that, The self-supervised contrast learning is introduced to perform efficient feature extraction and lightweight model design on the improved space non-cooperative target matching and recognition model, and an optimized space non-cooperative target matching and recognition model is formed.

9. A non-transitory computer-readable storage medium, comprising: The optimized space non-cooperative target matching and recognition model is trained on the constructed space debris dataset according to a designed loss function, and a trained space non-cooperative target matching and recognition model is obtained. The memory stores program instructions executed by the processor, and the processor calls the program instructions to execute the steps of the method for identifying lunar and earth space non-cooperative targets based on the improved deep learning model according to any one of claims 1 to 6. The non-transitory computer readable storage medium stores computer instructions, and the computer instructions enable the computer to execute the steps of the method for identifying lunar and earth space non-cooperative targets based on the improved deep learning model according to any one of claims 1 to 6.

Images