Space target detection model and method based on YOLOv8 improved model

By improving the YOLOv8 model, the GPF-SPPF, CAFusion and EAM-CARAFE modules are adopted, and the problems of insufficient intelligence level and slow speed of traditional detection systems are solved, high-precision detection of space debris and satellites are achieved, and space situational awareness capabilities are improved. and detection speed.

CN119091223AActive Publication Date: 2024-12-06HENAN UNIVERSITY OF TECHNOLOGY +1

Patent Information

Application Number
CN202411265407.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-10
Publication Date
2024-12-06
Estimated Expiration
2044-09-10

AI Technical Summary

Technical Problem

Traditional space object detection systems have insufficient intelligence and poor multi-task adaptability when detecting space debris, artificial satellites and meteorites. The two-stage detectors of deep learning algorithms are slow, making it difficult to meet the real-time detection needs.

Method used

Based on YOLOv8 improved model, the GPF-SPPF module is used to replace the original SPPF module, connect the convolution and attention fusion module CAFusion, and replace the upsampling module with the EAM-CARAFE module in the neck network, and set up multiple processing paths and volumes. Accumulate attention information module to extract local and global features and enhance feature extraction capabilities.

Benefits of technology

It significantly improves the detection accuracy of space debris and satellites, improves space situation awareness, reduces the amount of model parameters, reduces the calculation cost, and can be deployed on resource-constrained terminal devices, ensuring detection accuracy and speed.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119091223A_ABST
    Figure CN119091223A_ABST
Patent Text Reader

Abstract

The invention provides a space target detection model and method based on a YOLOv8 improved model, and the following improvements are made on an existing YOLOv8 model: in a backbone network, replacing an original SPPF module with a GPF-SPPF module; a convolution and attention fusion module CAFusion is connected to the rear part of the GPF-SPPF module; in a check neck network, an EAM-CARAFE module is used to replace an original up-sampling module. According to the model, space debris and space targets such as satellites of various models can be accurately identified, the space situation awareness capability is remarkably improved, and correct orbit maneuver decision-making of an on-orbit spacecraft is facilitated. Compared with a traditional target recognition algorithm, the method provided by the invention can extract the features of the space target in a richer manner by using a deep learning technology, thereby remarkably improving the detection accuracy of the space target.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of machine learning, and in particular to a space target detection model and method based on an improved YOLOv8 model. Background Art

[0002] With the rapid development of space technology, the number of orbiting satellites has increased significantly, and the importance of satellites in communication, navigation, monitoring and reconnaissance has become increasingly prominent. However, with the increase in the number of satellite launches, the number of space debris and abandoned satellites has also increased, seriously threatening space security. Therefore, countries have made the use of outer space resources a strategic development focus, and space target detection technology has become a hot research area.

[0003] Traditional space target detection systems are relatively mature in the precise positioning of natural celestial bodies such as stars and planets. However, when detecting near-Earth space targets such as space debris, artificial satellites and meteorites, traditional methods often perform poorly due to the complex motion trajectories, diverse appearances and complex lighting conditions of these targets. They have problems such as insufficient intelligence and poor multi-task adaptability. For example, the MF02A04 satellite-borne laser terminal in the "Jilin-1" constellation successfully conducted a satellite-to-ground laser high-speed image transmission experiment, laying the foundation for the transmission of satellite-to-ground optical detection data. Space-based optical detection systems have broad application prospects in the detection of space debris, abandoned satellites, meteorites, collision warnings and satellite integrity detection. For example, China's Gaofen-6 and Shijian-21 satellites have all-weather and all-round detection capabilities.

[0004] Traditional target detection methods mainly focus on feature extraction (enhancing feature expression and anti-distortion capabilities) and feature classification (improving classification accuracy and speed). Researchers have proposed a variety of feature extraction and classification methods, such as SIFT, Hough transform, AdaBoost, and DPM. However, these hard feature-based methods have the following three main problems: (a) insufficient target feature expression capabilities; (b) high feature similarity, resulting in a high classification error rate; (c) feature design is highly targeted, and a single feature is not effective in detecting multiple targets in complex scenes.

[0005] With the rise of convolutional neural networks (CNNs), deep learning has been widely used in the field of object detection and has gradually become a mainstream method. Researching deep learning-based space object detection algorithms is of great significance for spacecraft and satellites to autonomously identify surrounding objects and avoid space collisions. At present, deep learning object detection algorithms are mainly divided into two categories: one-stage and two-stage. Two-stage algorithms (such as Faster R-CNN) first generate candidate regions through a region proposal network (RPN), and then classify and regress these regions. One-stage algorithms (such as YOLO and SSD) complete object classification and position regression simultaneously in one forward propagation, so the detection speed is faster. Although two-stage detectors have advantages in detection accuracy, they are relatively slow because they require two feature extractions and calculations. Summary of the invention

[0006] The purpose of the present invention is to solve the deficiencies in existing research and provide a space target detection model and method based on the improved YOLOv8 model. The model of the present invention significantly improves the detection accuracy of space debris and satellites while ensuring the detection speed.

[0007] In the first aspect, the present invention provides a space target detection model based on the improved YOLOv8 model, and makes the following improvements to the existing YOLOv8 model:

[0008] In the backbone network, the original SPPF module is replaced with the GPF-SPPF module;

[0009] A convolution and attention fusion module CAFusion is connected after the GPF-SPPF module;

[0010] In the neck network, the EAM-CARAFE module is used to replace the original upsampling module.

[0011] Based on the above, the GPF-SPPF module sets three processing paths;

[0012] The main path includes three sequentially connected maximum pooling layers, which are used to generate three features containing local saliency information respectively;

[0013] The first sub-path passes through a global average pooling layer;

[0014] The second sub-path passes through a global maximum pooling layer;

[0015] The features generated by the three paths are concatenated, and then channel integration and further feature extraction are performed through the CBS convolutional layer to generate the final output features.

[0016] Based on the above, the convolution and attention fusion module CAFusion includes a first normalization module LayerNorm, a convolution attention information module CAIM, a second normalization module LayerNorm, and a multi-scale feedforward network module MFFN connected in sequence;

[0017] The convolutional attention information module CAIM is used to capture long-range dependencies and neighborhood spectral correlations, and the multi-scale feedforward network module MFFN is used to extract features of different scales and suppress multi-scale features;

[0018] The input of the second normalization module LayerNorm is also connected to the input of the first normalization module LayerNorm;

[0019] The output of the convolutional attention information module CAIM is also connected to the output of the multi-scale feedforward network module MFFN.

[0020] Based on the above, the convolutional attention information module CAIM includes a global branch and a local branch for calculating convolutional attention information;

[0021] The convolutional attention information of the local branch is expressed as:

[0022] F CONV =C 3×3×3 (C shuffle (C 1×1 (x)))

[0023] F CONV is the local branch output, C 1×1 is a 1×1 convolution, C 3×3×3 is a 3×3×3 convolution, C shuffle is the channel shuffle operation, x is the input feature;

[0024] In the global branch, we first generate the query Q, key K, and value value through a 1×1 convolution and three 3×3 depth convolutions, respectively, to generate three shapes: Tensor;

[0025] Next, reconstruct Q as K is reconstructed into According to and The interaction between them calculates the attention mechanism In order to reduce the calculation, the lightweight model is not calculated. Attention map;

[0026] Define the convolutional attention information of the global branch as F att :

[0027]

[0028] α is a learnable scaling parameter that controls the size of the matrix multiplication of K and Q before applying the softmax function; Y represents the original input, which is added to the output of the attention mechanism to form a residual connection, A representative value vector, representing the actual input data, which will be weighted according to the attention score;

[0029] Finally, the calculation result of the convolutional attention information module CAIM is: F out =F att +F conv .

[0030] Based on the above, the multi-scale feedforward network module MFFN uses two 1×1 convolutions to expand the feature channel into two feature channels, the input features are processed in two parallel feature channels, and a gating mechanism is introduced, and the element-by-element product of the two feature channels is used as the output to enhance the nonlinear transformation;

[0031] In one of the feature channels, the input features are processed by two parallel dilated convolution paths and then merged and output. In the other feature channel, the input features are processed by a 3×3 dilated convolution and then output.

[0032] The gating mechanism is introduced in another feature channel.

[0033] Based on the above, the EMA-CARAFE module includes a feature compression module, an upsampling kernel prediction module and a feature recombination module;

[0034] The feature compression module is used to perform 1×1 convolution on the feature map of input shape H×W×C to compress the number of channels and obtain H×W×C m The feature map of

[0035] The upsampling kernel prediction module is connected to the feature compression module and is used to perform H×W×C m The feature map of is used for content encoding and upsampling kernel prediction; among them, through K encoder ×K encoder The convolution layer predicts the upsampling kernel, and the number of input channels is C m , the output is σ 2 K up 2 , and expand the channel in the spatial dimension to obtain a shape of σH×σW×K up 2 The upsampling kernel of

[0036] The normalization module is connected to the upsampling kernel prediction module and is used to perform a normalization operation on the upsampling kernel so that the sum of the convolution kernel weights is 1;

[0037] The feature recombination module is connected to the normalization module and is used to perform the following operations:

[0038] Map each position in the normalized feature map back to the input feature map, and take K as the center of each position mapped back to the input feature map. up ×K up The area of ​​​​the normalized upsampling kernel is dot-producted to obtain the output value; different channels at the same position share the same upsampling kernel, and finally the output feature map of σH×σW×C is obtained;

[0039] H is the height of the input feature map, W is the width of the input feature map, and C is the number of channels of the input feature map; C m is the number of channels of the compressed feature map, K encoder ×K encoder represents the size of a convolution kernel used for upsampling kernel prediction, σ represents the upsampling ratio, K up 2 Indicates the size of the upsampling kernel.

[0040] In a second aspect, the present invention provides a space target detection method, comprising:

[0041] Step 1: Use STK software to simulate real space scenes, collect images of starry sky and near-target scenes, and create image datasets; use labeling tools to annotate images to form label datasets;

[0042] Step 2: Integrate the image dataset and the label dataset to make a complete training dataset; divide the dataset into a training set, a validation set, and a test set in a ratio of 8:1:1 to make a training dataset;

[0043] Step 3: Input the training data set into the space target detection model based on the improved YOLOv8 model for iterative training. When the performance index of the model on the validation set reaches the preset accuracy threshold, stop the training, save the parameter weights of the model at this time, perform starry sky near target detection on the test image data, and complete the model training;

[0044] Step 4: Input the image to be detected into the trained model, and the model outputs the detection result.

[0045] In a third aspect, the present invention provides a computer device, comprising:

[0046] one or more processors;

[0047] a memory for storing one or more programs,

[0048] When the one or more programs are executed by the one or more processors, the one or more processors perform the steps of the space target detection method as described.

[0049] In a fourth aspect, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the space target detection method as described above.

[0050] In a fifth aspect, the present invention provides a computer program product, comprising a computer program / instruction, characterized in that when the computer program / instruction is executed by a processor, the steps of the space target detection method as described above are implemented.

[0051] Compared with the prior art, the present invention has outstanding substantive features and significant progress, specifically:

[0052] 1. The model of the present invention can accurately identify space debris and space targets such as various types of satellites, significantly improving the space situation awareness capability and helping in-orbit spacecraft to make correct orbital maneuvering decisions.

[0053] 2. Compared with traditional target recognition algorithms, the present invention uses deep learning technology to extract more abundant features of space targets, thereby significantly improving the detection accuracy of space targets.

[0054] 3. The present invention effectively reduces the number of model parameters and computing costs by optimizing the model structure. Moreover, the model can be deployed on resource-constrained terminal devices, which not only ensures detection accuracy but also improves detection speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is the network structure of Context Sensing-YOLOv8 in Example 1 of the present invention.

[0056] Figure 2 This is a network structure diagram of the GPF-SPPF module in Example 1 of the present invention.

[0057] Figure 3 This is a network structure diagram of the CAFusion module in Example 1 of the present invention.

[0058] Figure 4 This is a structural diagram of the CAIM module in Example 1 of the present invention.

[0059] Figure 5 4 is a structural diagram of a multi-scale feedforward network MFFN in Example 1 of the present invention.

[0060] Figure 6 This is a structural diagram of the EAM-CARAFE in Example 1 of the present invention.

[0061] Figure 7 This is an evaluation index diagram in Example 1 of the present invention.

[0062] Figure 8 The following is a diagram showing the detection effects of the Context Sensing-YOLOv8 model and the original YOLOv8n model in the starry sky near-target scene. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0064] Example 1

[0065] This embodiment proposes a spatial target detection model based on the improved YOLOv8 model - Context Sensing-YOLOv8, such as Figure 1 Context Sensing-YOLOv8 is a model that makes the following improvements to the existing YOLOv8 model:

[0066] In the backbone network, the original SPPF module is replaced with the GPF-SPPF module;

[0067] A convolution and attention fusion module CAFusion is connected after the GPF-SPPF module;

[0068] In the neck network, the EAM-CARAFE module is used to replace the original upsampling module.

[0069] Specifically, Figure 2 As shown, the GPF-SPPF module sets three processing paths;

[0070] The main path includes three sequentially connected maximum pooling layers, which are used to generate three features containing local saliency information respectively;

[0071] The first sub-path passes through a global average pooling layer;

[0072] The second sub-path passes through a global maximum pooling layer;

[0073] The features generated by the three paths are concatenated, and then the CBS convolutional layer is used for channel integration and further feature extraction to generate the final output features;

[0074] The GPF-SPPF module of this embodiment can add some global background information and edge information to assist the feature extraction network to better extract target features.

[0075] In order to enhance the network's modeling capabilities for global and local features, this embodiment designs a convolution and attention fusion module CAFusion. Figure 3 As shown, CAFusion includes a first normalization module LayerNorm, a convolutional attention information module CAIM, a second normalization module LayerNorm, and a multi-scale feedforward network module MFFN connected in sequence;

[0076] The convolutional attention information module CAIM is used to capture long-range dependencies and neighborhood spectral correlations, and the multi-scale feedforward network module MFFN is used to extract features of different scales and suppress multi-scale features;

[0077] The input of the second normalization module LayerNorm is also connected to the input of the first normalization module LayerNorm;

[0078] The output of the convolutional attention information module CAIM is also connected to the output of the multi-scale feedforward network module MFFN.

[0079] Preferably, the convolutional attention information module CAIM includes a global branch and a local branch, such as Figure 4 As shown, it is used to calculate the convolutional attention information;

[0080] The convolutional attention information of the local branch is expressed as:

[0081] F CONV =C 3×3×3 (C shuffle (C 1×1 (x)))

[0082] F CONV is the local branch output, C 1×1 is a 1×1 convolution, C 3×3×3 is a 3×3×3 convolution, C shuffle is the channel shuffle operation, x is the input feature;

[0083] In the global branch, we first generate the query Q, key K, and value value through a 1×1 convolution and three 3×3 depth convolutions, respectively, to generate three shapes: Tensor;

[0084] Next, reconstruct Q as K is reconstructed into According to and The interaction between them calculates the attention mechanism In order to reduce the calculation, the lightweight model is not calculated. Attention map;

[0085] Define the convolutional attention information of the global branch as F att :

[0086]

[0087] α is a learnable scaling parameter that controls the size of the matrix multiplication of K and Q before applying the softmax function; Y represents the original input, which is added to the output of the attention mechanism to form a residual connection, A representative value vector, representing the actual input data, which will be weighted according to the attention score;

[0088] Finally, the calculation result of the convolutional attention information module CAIM is: F out =F att +F conv .

[0089] Preferably, the multi-scale feed-forward network module MFFN, such as Figure 5 As shown in the figure, two 1×1 convolutions are used to expand the feature channel into two feature channels, the input features are processed in two parallel feature channels, and a gating mechanism is introduced, and the element-by-element product of the two feature channels is used as the output to enhance the nonlinear transformation;

[0090] In one of the feature channels, the input features are processed by two parallel dilated convolution paths and then merged and output. In the other feature channel, the input features are processed by a 3×3 dilated convolution and then output. The gating mechanism is introduced in the other feature channel.

[0091] The formula for the gated output of the multi-scale feedforward network module MFFN is as follows:

[0092]

[0093] X out =C 1×1 Gating(X)

[0094] Among them, φ represents nonlinearity, e represents unit multiplication, represents a 3×3 dilated convolution with a dilation rate of 3. Represented as a 3×3 dilated convolution with a dilation rate of 2.

[0095] The traditional upsampling module ignores the semantic information of the feature image context, resulting in loss of detail information in the target-dense image. To solve this problem, this embodiment also provides an EMA-CARAFE module. Figure 6 As shown, the EMA-CARAFE module includes a feature compression module, an upsampling kernel prediction module and a feature recombination module;

[0096] The feature compression module is used to perform 1×1 convolution on the feature map of input shape H×W×C to compress the number of channels and obtain H×W×C m The feature map of

[0097] The upsampling kernel prediction module is connected to the feature compression module and is used to perform H×W×C m The feature map of is used for content encoding and upsampling kernel prediction; among them, through K encoder ×K encoder The convolution layer predicts the upsampling kernel, and the number of input channels is C m , the output is σ 2 K up 2 , and expand the channel in the spatial dimension to obtain a shape of σH×σW×K up 2 The upsampling kernel of

[0098] The normalization module is connected to the upsampling kernel prediction module and is used to perform a normalization operation on the upsampling kernel so that the sum of the convolution kernel weights is 1;

[0099] The feature recombination module is connected to the normalization module and is used to perform the following operations:

[0100] Map each position in the normalized feature map back to the input feature map, and take K as the center of each position mapped back to the input feature map. up ×K up The area of ​​​​the normalized upsampling kernel is dot-producted to obtain the output value; different channels at the same position share the same upsampling kernel, and finally the output feature map of σH×σW×C is obtained;

[0101] H is the height of the input feature map, W is the width of the input feature map, and C is the number of channels of the input feature map (also called depth);

[0102] C m is the number of channels of the compressed feature map. In the upsampling prediction module, the input feature map is compressed to C after 1×1 convolution. m , in order to reduce the amount of subsequent calculations;

[0103] K encoder ×K encoder Represents the size of a convolution kernel used for upsampling kernel prediction;

[0104] σH×σW: This is the size of the output feature map after upsampling, σ 2 K up 2 represents the number of channels of the upsampling kernel, where σ represents the upsampling ratio, K up 2Indicates the size of the upsampling kernel.

[0105] The output formula of the EMA-CARAFE module is as follows:

[0106] W i =ψ(N(x i ,k encode ),x i ·σ(a i ))

[0107] Where: ψ is the upsampling prediction module, σ is the sigmoid activation function, N(x i ,k) represents the k*k area centered at position i in the input feature map x;

[0108] The result of the upsampling kernel prediction module is multiplied by the result of the feature reorganization module to obtain the upsampling result corresponding to the i position in the input feature map x in the output result map:

[0109]

[0110] Where: It is the feature recombination module.

[0111] In the target detection algorithm, recall rate, precision, mAP and inference speed are usually selected as evaluation indicators of the network model. Recall rate P represents the proportion of correctly identified positive classes to the total positive classes, and precision R reflects the proportion of positive classes correctly classified by the model;

[0112]

[0113] mAP integrates the Precision and Recall of all categories by calculating the area under the PR curve of each category. The higher the value, the better the multi-category performance of the model.

[0114]

[0115] Evaluation indicators such as Figure 7 As shown, the model proposed in this embodiment improves the mAP value of the original model by 0.8%, P by 0.8%, and R by 0.7%. Since the detection speed is 108 frames per second, which is greater than 25 frames per second, it meets the requirements of real-time monitoring.

[0116] Example 2

[0117] This embodiment provides a space target detection method, including:

[0118] Step 1: Use STK software to simulate real space scenes, collect images of starry sky and near-target scenes, and create image datasets; use labeling tools to annotate images to form label datasets, such as Figure 7 As shown;

[0119] Step 2: Integrate the image dataset and the label dataset to make a complete training dataset; divide the dataset into a training set, a validation set, and a test set in a ratio of 8:1:1 to make a training dataset;

[0120] Step 3: Input the training data set into the space target detection model based on the improved YOLOv8 model described in Example 1 for iterative training, stop training when the performance index of the model on the validation set reaches a preset accuracy threshold, save the parameter weights of the model at this time, perform starry sky near target detection on the test image data, and complete model training;

[0121] Step 4: Input the image to be detected into the trained model, and the model outputs the detection result.

[0122] Figure 8 The following is a diagram of the detection effects of the Context Sensing-YOLOv8 model and the original YOLOv8n model in the starry sky near-target scene. It can be seen that the accuracy of Context Sensing-YOLOv8 in space target detection is greatly improved compared with the original YOLOv8n model.

[0123] Example 3

[0124] Each module in the method of embodiment 2 can be implemented in whole or in part by software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute operations corresponding to each module.

[0125] In an exemplary embodiment, a computer device is provided, which may be a terminal. The computer device further includes a processor, a memory, an input / output interface, a communication interface, a display unit, and an input device. The processor, the memory, and the input / output interface are connected via a system bus, and the communication interface, the display unit, and the input device are connected to the system bus via the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be implemented through WIFI, a mobile cellular network, NFC (near field communication) or other technologies. When the computer program is executed by the processor, the steps of the space target detection method described in Example 2 are implemented. The display unit of the computer device is used to form a visually visible picture, which may be a display screen, a projection device, or a virtual reality imaging device. The display screen can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad set on the computer device shell, or an external keyboard, touchpad or mouse.

[0126] Those skilled in the art will appreciate that the structure of the above-mentioned computer device is only a partial structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components, or combine certain components, or have a different arrangement of components.

[0127] In an exemplary embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the space target detection method described in Example 2 are implemented.

[0128] In an exemplary embodiment, a computer program product is provided, including a computer program / instruction, which implements the steps of the space target detection method as described in Example 2 when executed by a processor.

[0129] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0130] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0131] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A space target detection model based on the improved YOLOv8 model, characterized in that: The following improvements are made to the existing YOLOv8 model: In the backbone network, the original SPPF module is replaced with the GPF-SPPF module; A convolution and attention fusion module CAFusion is connected after the GPF-SPPF module; In the neck network, the EAM-CARAFE module is used to replace the original upsampling module.

2. The space target detection model based on the improved YOLOv8 model according to claim 1, characterized in that: The GPF-SPPF module sets three processing paths; The main path includes three sequentially connected maximum pooling layers, which are used to generate three features containing local saliency information respectively; The first sub-path passes through a global average pooling layer; The second sub-path passes through a global maximum pooling layer; The features generated by the three paths are concatenated, and then channel integration and further feature extraction are performed through the CBS convolutional layer to generate the final output features.

3. The space target detection model based on the improved YOLOv8 model according to claim 1, characterized in that: The convolution and attention fusion module CAFusion includes a first normalization module LayerNorm, a convolution attention information module CAIM, a second normalization module LayerNorm, and a multi-scale feedforward network module MFFN connected in sequence; The convolutional attention information module CAIM is used to capture long-range dependencies and neighborhood spectral correlations, and the multi-scale feedforward network module MFFN is used to extract features of different scales and suppress multi-scale features; The input of the second normalization module LayerNorm is also connected to the input of the first normalization module LayerNorm; The output of the convolutional attention information module CAIM is also connected to the output of the multi-scale feedforward network module MFFN.

4. The space target detection model based on the improved YOLOv8 model according to claim 3, characterized in that: The convolutional attention information module CAIM includes a global branch and a local branch, which are used to calculate the convolutional attention information; The convolutional attention information of the local branch is expressed as: F CONV =C 3×3×3 (C shuffle (C 1×1 (x))) F CONV is the local branch output, C 1×1 is a 1×1 convolution, C 3×3×3 is a 3×3×3 convolution, C shuffle is the channel shuffle operation, x is the input feature; In the global branch, we first generate the query Q, key K, and value value through a 1×1 convolution and three 3×3 depth convolutions, respectively, to generate three shapes: Tensor; Next, reconstruct Q as K is reconstructed into According to and The interaction between them calculates the attention mechanism In order to reduce the calculation, the lightweight model is not calculated. Attention map; Define the convolutional attention information of the global branch as F att : α is a learnable scaling parameter that controls the size of the matrix multiplication of K and Q before applying the softmax function; Y represents the original input, which is added to the output of the attention mechanism to form a residual connection, A representative value vector, representing the actual input data, which will be weighted according to the attention score; Finally, the calculation result of the convolutional attention information module CAIM is: F out =F att +F conv .

5. The space target detection model based on the improved YOLOv8 model according to claim 3, characterized in that: The multi-scale feedforward network module MFFN uses two 1×1 convolutions to expand the feature channel into two feature channels. The input features are processed in two parallel feature channels, and a gating mechanism is introduced. The element-by-element product of the two feature channels is used as the output to enhance the nonlinear transformation. In one of the feature channels, the input features are processed by two parallel dilated convolution paths and then merged and output. In the other feature channel, the input features are processed by a 3×3 dilated convolution and then output. The gating mechanism is introduced in another feature channel.

6. The space target detection model based on the improved YOLOv8 model according to claim 1, characterized in that: The EMA-CARAFE module includes a feature compression module, an upsampling kernel prediction module and a feature recombination module; The feature compression module is used to perform 1×1 convolution on the feature map of input shape H×W×C to compress the number of channels and obtain H×W×C m The feature map of The upsampling kernel prediction module is connected to the feature compression module and is used to perform H×W×C m The feature map of is used for content encoding and upsampling kernel prediction; among them, through K encoder ×K encoder The convolution layer predicts the upsampling kernel, and the number of input channels is C m , the output is And expand the channel in the spatial dimension to get the shape The upsampling kernel of The normalization module is connected to the upsampling kernel prediction module and is used to perform a normalization operation on the upsampling kernel so that the sum of the convolution kernel weights is 1; The feature recombination module is connected to the normalization module and is used to perform the following operations: Map each position in the normalized feature map back to the input feature map, and take K as the center of each position mapped back to the input feature map. up ×K up The area of ​​​​the normalized upsampling kernel is dot-producted to obtain the output value; different channels at the same position share the same upsampling kernel, and finally the output feature map of σH×σW×C is obtained; H is the height of the input feature map, W is the width of the input feature map, and C is the number of channels of the input feature map; C m is the number of channels of the compressed feature map, K encoder ×K encoder represents the size of a convolution kernel used for upsampling kernel prediction, σ represents the upsampling ratio, Indicates the size of the upsampling kernel.

7. A space target detection method, characterized in that: include: Step 1: Use STK software to simulate real space scenes, collect images of starry sky and near-target scenes, and create image datasets; use labeling tools to annotate images to form label datasets; Step 2: Integrate the image dataset and the label dataset to make a complete training dataset; divide the dataset into a training set, a validation set, and a test set in a ratio of 8:1:1 to make a training dataset; Step 3: Input the training data set into the space target detection model based on the improved YOLOv8 model described in any one of claims 1 to 6 for iterative training, stop the training when the performance index of the model on the validation set reaches a preset accuracy threshold, save the parameter weights of the model at this time, perform starry sky near target detection on the test image data, and complete the model training; Step 4: Input the image to be detected into the trained model, and the model outputs the detection result.

8. A computer device, characterized in that: include: one or more processors; a memory for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors perform the steps of the space target detection method as described in claim 7.

9. A computer-readable storage medium storing a computer program, characterized in that: When the program is executed by a processor, the steps of the space target detection method as claimed in claim 7 are implemented.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the space target detection method as claimed in claim 7 are implemented.

Citation Information

Patent Citations

  • Traffic sign detection algorithm based on improved YOLOV5

    CN117218335A

  • Lightweight real-time strip steel surface defect detection method based on improved YOLOv5s model

    CN118505658A

  • Small target detection method, system and equipment based on improved YOLOv8

    CN118552935A

  • Synthetic aperture radar (SAR) image target detection method

    US20230169623A1

  • Expression recognition method based on attention-modulated contextual spatial information

    WO2023185243A1

Cited By

  • On-orbit moving target detection method in inertial space stable imaging mode

    CN120070555A