Target detection method and system based on superscanning electroencephalogram sharing space feature guidance

Abstract

The invention discloses a target detection method and system based on superscanning electroencephalogram sharing space feature guidance. According to the target detection method, shared channel pairs are screened from channels of a plurality of private spaces according to the magnitude of synchronism, and a shared space is constructed. And respectively enhancing each private feature by using the shared space feature to obtain enhanced electroencephalogram signal features corresponding to a plurality of private spaces. And fusing to obtain a final classification feature. And inputting the final classification features into a classifier to obtain a target detection result of the detected image. According to the method, the common cognitive features between the two subjects are captured through the shared space, and the individual specific features are extracted in combination with the private space, so that the accuracy and efficiency of target detection are effectively improved. In addition, a multi-head attention mechanism in the joint learning framework further optimizes information interaction between sharing and private spaces, so that the model has stronger generalization ability and robustness in multiple task scenes.

Description

Technical Field

[0001] The present invention belongs to the technical field of brain-computer collaborative target detection, and specifically relates to a target detection method and system based on hyperscanning-based EEG shared spatial feature guidance. Background Art

[0002] A brain-computer interface (BCI) system is a technology that enables direct communication between the human brain and external devices. It is widely used in fields such as medical rehabilitation, intelligent control, and object detection. Electroencephalography (EEG)-based BCIs have become a research hotspot due to their non-invasiveness, ease of use, and high temporal resolution.

[0003] Rapid Serial Visual Presentation (RSVP) is a commonly used paradigm for EEG-based object detection tasks. RSVP utilizes event-related potential (ERP) features by presenting stimulus images in rapid succession for object detection. Typical ERP components include the P300 and N200, which produce distinct EEG responses to specific stimuli. However, due to issues such as low signal-to-noise ratio, nonlinearity, and individual variability in EEG signals, the detection performance of traditional single-brain detection systems in complex scenes remains limited.

[0004] In recent years, hyperscanning, a novel neuroimaging method, has attracted widespread attention. Hyperscanning simultaneously records EEG signals from multiple subjects, allowing for the investigation of cross-brain synchronization and information integration during collaborative tasks. This technique has been shown to be useful for exploring the neural mechanisms underlying collective decision-making and team interaction. However, most current studies are limited to simple data concatenation or statistical analysis, lacking deep feature fusion and modeling approaches guided by synchronization information.

[0005] As research continues to deepen, existing methods still face limitations. First, existing RSVP systems primarily rely on single-brain signal processing, making it difficult to fully utilize the synchronous information in cross-brain collaboration, thus limiting their ability to analyze complex tasks. Second, while multi-trial methods can improve signal stability, they require longer experimental times and computational complexity, making real-time applications challenging. Furthermore, current research on cross-brain collaboration has largely focused on simple data concatenation or statistical analysis, lacking deep feature fusion and modeling mechanisms guided by synchronous information, hindering the full potential of cross-brain collaboration. Therefore, developing methods that can efficiently utilize cross-brain information and possess real-time processing capabilities has become a research priority. Furthermore, traditional EEG decoding methods typically rely on individual EEG activity data and rarely consider cross-brain coupling information. However, hyperscanning can simultaneously record EEG signals from multiple participants, thereby capturing cross-brain synchronization in collaborative tasks and improving decision-making reliability. Therefore, a new decoding method is needed that can simultaneously consider both shared and private features to improve object detection performance in multi-brain collaboration. Summary of the Invention

[0006] The purpose of the present invention is to address the deficiencies of the existing technology and provide a target detection method and system based on hyperscanning-based EEG shared space feature guidance. It designs a dual-subject RSVP paradigm under the same task, synchronously collects dual-brain EEG signals through hyperscanning technology, constructs a shared space by calculating the phase synchronization value (PLV), and combines multi-scale convolutional networks and multi-head attention mechanisms to extract and fuse features to achieve more efficient and accurate target detection. This method can dynamically adjust the subject's private feature weights, extract instructive mutual information from the shared space, and enhance the EEG feature expression ability of each subject.

[0007] In a first aspect, the present invention provides a target detection method guided by EEG shared spatial features based on hyperscanning; the method comprises: The EEG characteristics of multiple observers when observing the measured image are used as multiple private spaces.

[0008] Based on the degree of synchronization, one or more shared channel pairs are selected from each channel in multiple private spaces. Each shared channel pair includes channels at the same location in different private spaces. All shared channel pairs constitute a shared space.

[0009] Feature extraction is performed on the shared space and each private space respectively to obtain shared space features and multiple private features.

[0010] The shared space features are used to enhance each private feature respectively, and the enhanced EEG signal features corresponding to multiple private spaces are obtained.

[0011] The features of each enhanced EEG signal are fused to obtain the final classification features.

[0012] The final classification features are input into the classifier to obtain the target detection results.

[0013] Preferably, the process of selecting the shared channel pairs is as follows: calculating phase lock values ​​for corresponding channels in different private spaces, and extracting channels in each private space in descending order of phase lock values ​​as shared channel pairs.

[0014] Preferably, the feature extraction comprises a single or serially connected multi-scale convolution operation. The multi-scale convolution operation comprises multiple convolution layers with different convolution kernels. Each convolution layer comprises a temporal convolution block and a fine depth convolution block. The convolution layers within the same multi-scale convolution module are arranged in parallel, and the output features are concatenated. Batch normalization, dropout, and activation function processing are performed after each convolution block.

[0015] Preferably, the process of enhancing each private feature with shared spatial features is as follows: based on the multi-head cross-attention mechanism, the shared spatial features are linearly converted into a query matrix; the private features are linearly converted into a key matrix and a value matrix; the attention scores in each attention head are calculated and concatenated respectively to obtain enhanced EEG signal features.

[0016] Preferably, multiple enhanced EEG signal features are fused by decision making to obtain final classification features.

[0017] Preferably, the number of the observers is two. The classifier is trained using a data set including EEG signals of each subject when observing an image, and a loss function adopts a cross entropy loss.

[0018] Preferably, the multiple private spaces are obtained by performing superscanning synchronous EEG acquisition on multiple observers who simultaneously observe the measured image, and preprocessing the acquired EEG signals.

[0019] In a second aspect, the present invention provides a target detection system based on hyperscanning EEG shared space feature guidance, which is used to perform the aforementioned target detection method. The target detection system includes an EEG extraction module, a shared space construction module, a feature extraction module, a feature guidance fusion module and a classifier. The EEG extraction module is used to collect and preprocess the EEG features of the observer when observing the measured image; the shared space construction module is used to establish a shared space based on each private space. The feature extraction module is used to extract features from EEG features, including a single layer or a multi-layer multi-scale convolution module connected in series. The multi-scale convolution module includes multiple convolution layers that are set in parallel and output feature splicing. Each convolution layer includes a temporal convolution block and a fine depth convolution block. The convolution kernel size of each convolution layer is different.

[0020] The feature-guided fusion module includes multiple fusion enhancement branches corresponding to the number of private features. Each private feature is fed into its corresponding fusion enhancement branch. Shared spatial features are fed into all fusion enhancement branches simultaneously. A multi-head cross-attention mechanism is employed in both fusion enhancement branches. The output features of each fusion enhancement branch are fused and then output to the classifier.

[0021] In a third aspect, the present invention provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the memory stores the computer program; and the processor executes the aforementioned target detection method.

[0022] In a fourth aspect, the present invention provides a readable storage medium storing a computer program; when the computer program is executed by a processor, it is used to implement the aforementioned target detection method.

[0023] The present invention has the following beneficial effects: 1. This paper proposes a novel joint learning method that seamlessly integrates shared space representation with private space learning. This method captures the common cognitive features between two observers by using the shared space, while extracting individual-specific features in the private space, effectively improving the accuracy and efficiency of object detection. Furthermore, the multi-head attention mechanism in the joint learning framework further optimizes the information exchange between the shared and private spaces, resulting in a model with stronger generalization and robustness across a variety of task scenarios.

[0024] 2. This paper uses quantified synchronization to construct a shared space based on the EEG features of multiple observers, selecting the most collaborative channel pairs. This approach not only accurately captures inter-subject collaboration but also integrates collaborative features into individual features by guiding the learning of private space representations, enhancing the model's adaptability to complex tasks. Furthermore, the separation of shared space features from private features makes the model's decision-making process more transparent and explainable, providing strong support for further understanding of multi-brain collaboration.

[0025] 3. The present invention uses hyperscanning technology to synchronously collect EEG signal data of two subjects in a collaborative target recognition task, generating a high-quality data set containing target and non-target stimuli. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 is a flow chart of an embodiment of the present invention; Figure 2 Schematic diagram of the experimental paradigm for EEG data collection in Example 1 of the present invention.

[0027] Figure 3 This is a schematic diagram of the framework for shared space construction, feature extraction, and feature guidance fusion in Example 1 of the present invention.

[0028] Figure 4 Schematic diagram of the feature extraction architecture in Example 1 of the present invention.

[0029] Figure 5 Schematic diagram of the architecture of shared spatial feature guided fusion in Example 1 of the present invention. DETAILED DESCRIPTION

[0030] The method of the present invention is described in detail below with reference to the accompanying drawings.

[0031] Example 1 like Figure 1 As shown, a target detection method guided by EEG shared spatial features based on hyperscanning includes the following steps: Step 1: EEG data collection This example uses renderings from a simulated real-world environment as stimulus material. All images are from a custom-made simulation. Targets are inserted in different sequences and may appear in various positions. To align with real-world conditions, the target's outline and color are roughly similar to those of the background (making them difficult to distinguish). Non-target objects, such as vehicles, are also present for interference.

[0032] EEG acquisition is carried out in a closed, light-free environment to ensure that the data acquisition process is not affected by external light and noise. This embodiment involves two subjects, and high-precision synchronous acquisition of EEG signals is achieved through superscanning technology. This technology ensures that the neural activities of the two subjects can be compared and analyzed in real time during the experiment. The two subjects face the same model of screen, the distance between the screen and the subject is 75cm, and the screen resolution is 1920×1080, ensuring consistent display effects and providing the same visual stimulation for the two subjects. Before data acquisition, the synchronization and quality of signal acquisition of the two subjects are ensured.

[0033] In this embodiment, data collection is divided into 8 sections, each section contains 10 groups, and each group consists of 4 trials (e.g. Figure 2 (As shown). During each trial, a video sequence containing 15 labeled images played on the screen: 13 non-target images and 2 target images, with a target-to-non-target ratio of approximately 1:6. Each image was displayed for 0.8 seconds, and a "+" symbol appeared in the center of the screen for 5 seconds between trials to help participants focus and adjust their state. The experimental design emphasized that two participants viewed identical visual stimuli for the same period of time, and hyperscanning technology was used to record their collaborative cognitive characteristics during this process.

[0034] To ensure high-quality data acquisition, the experiment used a 64-lead Neuroscan device to record EEG signals. Electrode placement followed the international 10-20 system, and the sampling frequency was set to 1000 Hz. Electrode impedance was kept below 15 kΩ to ensure stable and clear signals. During data acquisition, both subjects' EEG signals were monitored in real time to avoid data loss or poor quality. Furthermore, the two subjects used identical experimental conditions, including screen, chair height, and ambient lighting, to ensure a highly consistent and controllable experimental design.

[0035] Step 2: Data preprocessing The collected EEG data was filtered using a 0.1-40 Hz Butterworth filter to remove oculoculographic and myoelectrical interference. The data was then downsampled to 250 Hz, and the MNE library was used to remove and reconstruct damaged electrode lead channels to ensure signal integrity and accuracy, eliminating the impact of abnormal signals on subsequent analysis. This resulted in valid 62-electrode lead data. For each image, the EEG sample data for the subject was captured from the image's onset to 1.1 seconds after its onset. This yielded 4800 sample data points for each subject, including 640 target samples and 4160 non-target samples.

[0036] Step 3: Building a shared space like Figure 3 As shown in Stage 1 of the paper, the establishment of a shared space is one of the key steps in this embodiment. It captures the synchronization information of the two brain participants in the task and constructs a space that integrates shared and private features to improve the performance of the model in multi-brain cooperation tasks. In this stage, EEG signal data at the same time point are extracted from two subjects (i.e., two EEG data sources) as model input. The specific steps are as follows: The EEG signal data of the two subjects are processed into time series matrices, which are recorded as two private spaces. and The dimensions of each matrix are × ,in represents the number of EEG channels, For each subject, the signal data will be sampled synchronously in time to ensure that the two signals are processed at the same time.

[0037] In order to construct the shared space, the private spaces of the two subjects were first and Cross-brain coupling analysis was performed, specifically by calculating the phase synchronization quantification method. The phase lock value (PLV) was used to measure the synchronization between the EEG activities of two subjects on the time axis. The calculation formula of the phase lock value (PLV) is as follows:

[0038] in, and Respectively represent Channel at time The instantaneous phase of is the total number of time samples.

[0039] The higher the phase lock value PLV, the stronger the synchronization between the two private spaces. Synchronicity reflects the information synergy of the two subjects on the same channel. After completing the cross-brain synchronization analysis, one or more channel pairs with the strongest synchronization are selected as shared channel pairs. and The shared channel pair can best reflect the collaborative cognition between the two subjects.

[0040] Each shared channel pair constitutes a shared space, denoted as In this embodiment, five shared channel pairs are extracted to obtain a ten-channel shared space. The EEG signals in the shared feature space are believed to carry strong commonality information, reflecting the collaborative cognition of the two subjects in the task.

[0041] Create a shared space The expression is as follows:

[0042] in, k is the sequence number of the shared channel pair; is the number of shared channel pairs.

[0043] Step 4: Feature extraction like Figure 3 As shown in stage 2, the goal of feature extraction is to extract and , and the shared space built in step 3 Extract effective spatiotemporal features in order to facilitate the subsequent feature interaction and classifier learning. This process is completed by the feature extractor.

[0044] Given that EEG data has strong temporal characteristics, effectively capturing short-term and long-term temporal dependencies is crucial for preserving key neural dynamics. At the same time, EEG signals exhibit intricate spatial relationships between different channels, reflecting the interactions between different brain regions. Traditional methods often find it difficult to jointly model these temporal and spatial patterns; the feature extractor in this embodiment uses a multi-level, multi-scale convolutional neural network, whose network structure is as follows: Figure 4 The convolutional neural network includes two layers of multi-scale convolution modules connected in series.

[0045] The multi-scale convolution module includes three convolutional layers with different convolution kernels. Each convolutional layer includes a temporal convolution block and a fine depth convolution block. The three convolutional layers in the same multi-scale convolution module are set up in parallel, and the output features are spliced ​​together and sent to the next multi-scale convolution module or output externally.

[0046] In this embodiment, the convolutional neural network uses temporal convolution blocks and fine depth convolution blocks to extract rich features from EEG signals, facilitating subsequent feature interaction and classifier learning. The convolution kernel size is set to different scales to mine EEG signal features at different scales.

[0047] After each convolution block, batch normalization (BN), dropout, and activation function (ELU) are set. These operations help improve the training stability of the model and avoid overfitting.

[0048] The private features extracted from the two private spaces in this step and shared space features extracted from the shared space The expression is as follows:

[0049] in, is the feature extractor.

[0050] feature 、 Both are high-dimensional feature vectors containing multiple layers of features; shared spatial features It can effectively represent the neural synchronization signals of the high consensus parts of two subjects in the dual-brain collaboration task.

[0051] Step 5: Feature-guided fusion like Figure 3 As shown in stage 3, in this embodiment, a multi-head cross attention mechanism is adopted based on the shared spatial features extracted in step 4. , respectively optimize and enhance the private features of each subject , to achieve the effective fusion of shared spatial features and private features, so as to improve the model's ability to capture key information in dual-brain collaboration tasks; two private features Enhanced EEG signal features obtained by fusing shared spatial features The final classification features are obtained by fusion.

[0052] The specific implementation process is as follows: 5-1. Establish a feature-guided fusion model. Figure 5 As shown in Figure 2, the guidance fusion model includes two fusion enhancement branches and a decision fusion structure. Input two fusion enhancement branches respectively. Shared spatial features The two fusion enhancement branches are input at the same time. Multi-head cross attention mechanism is adopted in both fusion enhancement branches.

[0053] 5-2. Shared space features Linear transformation to query matrix , which is transformed by multiplying the learnable weight matrix Completed, the formula is expressed as:

[0054] in, h is the sequence number of the attention head in the multi-head cross attention mechanism.

[0055] At the same time, each private feature All through the corresponding weight matrix 、 Get the corresponding bond matrix and value matrix :

[0056] 5-3. In the two fusion enhancement branches, the attention scores of each attention head are calculated separately and concatenated to obtain enhanced EEG signal features. as follows:

[0057] 5-4. Enhanced EEG signal features output by two fusion enhancement branches Perform decision fusion to obtain the final classification features.

[0058] Step 6: Classifier learning The final classification features are input into the classifier. The classifier uses a multi-layer fully connected neural network. The classifier further extracts high-level features by calculating the distance between each feature vector in the vector space. Specifically, the classifier uses a distance metric based on the similarity of these feature vectors to evaluate the relevance of each input feature and the target category. Subsequently, these distances are processed by the Softmax function to generate a probability distribution for each category. The Softmax function converts the calculated distance into a standardized probability value, which represents the probability that the input data belongs to the target category. Ultimately, based on this probability distribution, the classifier can determine whether the image belongs to the target category and output the corresponding prediction result.

[0059] In the design and training of the classifier, the final classification features obtained in step 5 are input into a classifier for target classification. The classifier optimizes the target detection performance by adaptively adjusting the network weights. During the training process, this embodiment uses the batch gradient descent method to optimize the network, and uses the cross entropy loss function to measure the difference between the predicted results and the actual labels. In order to accelerate convergence and improve training stability, the Adam optimizer is introduced, which can adaptively adjust the learning rate according to the gradient of each parameter. In addition, in order to avoid overfitting, data augmentation and Dropout techniques are used in training, and the performance of the validation set is monitored by the early stopping method to prevent performance degradation caused by excessive training. Through these training strategies, the classifier can effectively integrate shared spatial features and private features to improve the accuracy and robustness of target detection.

[0060] In some other embodiments, the number of subjects is greater than 2. z , corresponding to the formation z private spaces; each shared channel pair generated in step 3 contains n channels. The feature-guided fusion model described in step 5 contains z fusion enhancement branches. z The values ​​of are, for example, 3, 4, 5, 6, and 7.

[0061] In some embodiments, the detected object is a specific vehicle, such as a car or a truck.

[0062] The effectiveness of the target detection method provided in this embodiment is verified by collecting data through specific experiments. The process is as follows: Considering the significant disparity in the number of targets and non-targets, this example uses the Balance Accuracy (BA) metric for detection results. This metric, designed for handling imbalanced datasets, measures the overall performance of a classifier by considering the detection performance of both positive and negative classes. Unlike traditional accuracy, BA provides a more balanced representation of classifier performance when the class population is imbalanced. BA calculates the accuracy for both the target class (positive class) and the non-target class (negative class) and averages the two. Specifically, BA calculates the recall for each class and then averages the two, thereby eliminating the impact of imbalanced datasets on the evaluation results.

[0063] When the ratio of target and non-target samples is severely unbalanced, traditional accuracy may be affected by the excessive number of non-target samples, resulting in a higher accuracy but actually poor recognition of the target class. By using BA, we can better measure the model's performance on both target and non-target classes, thereby obtaining more reliable evaluation results.

[0064]

[0065] Among them, TP represents true positive examples, which are the number of samples correctly classified as targets; TN represents true negative examples, which are the number of samples correctly classified as non-targets; FP represents false positive examples, which are the number of samples incorrectly classified as targets; and FN represents false negative examples, which are the number of samples incorrectly classified as non-targets.

[0066] The following table compares the detection results of this embodiment with those using other network models:

[0067] Experimental results show that the shared-space feature-guided fusion based on the dual-brain paradigm in this embodiment exhibits stronger cross-brain synchronization in the target detection task, especially in the event-related potential (ERP) analysis of the P300 component. These experimental results support the hypothesis that combining shared-space and private-space learning can enhance the robustness and accuracy of multi-participant brain-computer interface (BCI) systems in target detection.

[0068] Example 2 A target detection system includes an EEG extraction module, a shared space construction module, a feature extraction module, a feature-guided fusion module, and a classifier. The EEG extraction module is used to collect and preprocess the EEG features of an observer observing a measured image; the shared space construction module is used to establish a shared space based on each private space. The feature extraction module is used to extract EEG features and includes a single or multiple multi-scale convolution module connected in series. The multi-scale convolution module includes multiple convolution layers arranged in parallel and whose output features are spliced. Each convolution layer includes a temporal convolution block and a fine depth convolution block. The convolution kernels of each convolution layer vary in size. The feature-guided fusion module includes multiple fusion enhancement branches corresponding to the number of private features. Each private feature is input into a corresponding fusion enhancement branch. Shared space features are simultaneously input into all fusion enhancement branches. A multi-head cross-attention mechanism is employed in both fusion enhancement branches. The output features of each fusion enhancement branch are fused and output to the classifier.

[0069] Example 3 An embodiment of the present invention provides an electronic device. Specifically, the electronic device includes a memory and a processor. The memory stores executable code. When the processor executes the executable code, the method described in any one of the embodiments is implemented.

[0070] The memory may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage. The system network element communicates with at least one other network element via at least one communication interface (which may be wired or wireless), such as the Internet, a wide area network (WAN), a local area network (LAN), or a metropolitan area network (MAN).

[0071] The bus may be an ISA bus, a PCI bus or an EISA bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc.

[0072] The memory is used to store the program, and the processor executes the program after receiving the execution instruction. The method executed by the device for flow process definition disclosed in any of the aforementioned embodiments of the present invention can be applied to the processor or implemented by the processor.

[0073] The processor may be an integrated circuit chip with signal processing capabilities. During implementation, the steps of the above method may be completed by hardware integrated logic circuits in the processor or by software instructions. The above processor may be a general-purpose processor that can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of the present invention.

Claims

1. A target detection method based on hyperscanning EEG shared spatial feature guidance; characterized by: include: The EEG characteristics of multiple observers observing the measured image are used as multiple private spaces; According to the degree of synchronization, one or more shared channel pairs are selected from each channel of the multiple private spaces; all shared channel pairs constitute a shared space; Extract features of the shared space and each private space separately to obtain shared space features and multiple private features; Use the shared space features to enhance each private feature separately, and obtain enhanced EEG signal features corresponding to multiple private spaces; The enhanced EEG signal features are fused to obtain the final classification features; The final classification features are input into the classifier to obtain the target detection results.

2. The target detection method according to claim 1, wherein: The process of screening shared channel pairs is as follows: calculating the phase locking value for each corresponding channel in different private spaces respectively; extracting channels in each private space according to the phase locking value from large to small as shared channel pairs.

3. The target detection method according to claim 1, wherein: The feature extraction includes one layer or multiple layers of multi-scale convolution operations connected in series; the multi-scale convolution operation includes multiple convolution layers with different convolution kernels; each convolution layer includes a temporal convolution block and a fine depth convolution block; the convolution layers in the same multi-scale convolution module are set in parallel, and the output features are spliced ​​together.

4. The target detection method according to claim 1, wherein: The process of enhancing each private feature with shared spatial features is as follows: based on the multi-head cross-attention mechanism, the shared spatial features are linearly converted into a query matrix; the private features are linearly converted into a key matrix and a value matrix; the attention scores in each attention head are calculated and concatenated to obtain enhanced EEG signal features.

5. The target detection method according to claim 1, wherein: Multiple enhanced EEG signal features are fused through decision making to obtain the final classification features.

6. The target detection method according to claim 1, wherein: The number of observers is two; the classifier is trained using a data set including EEG signals of each subject when observing an image, and a loss function adopts a cross entropy loss.

7. The target detection method according to claim 1, wherein: The multiple private spaces are obtained by performing hyperscanning synchronous EEG acquisition on multiple observers who observe the measured image at the same time and preprocessing the acquired EEG signals.

8. A target detection system guided by hyperscanning EEG shared spatial features, characterized by: Used to execute a target detection method according to claim 1; the target detection system includes an EEG extraction module, a shared space construction module, a feature extraction module, a feature guidance fusion module and a classifier; the EEG extraction module is used to collect and preprocess the EEG features of the observer when observing the measured image; the shared space construction module is used to establish a shared space based on each private space; the feature extraction module is used to extract features from EEG features, including a single layer or a multi-layer multi-scale convolution module connected in series; the multi-scale convolution module includes multiple convolution layers set in parallel and outputting feature splicing; each convolution layer includes a temporal convolution block and a fine depth convolution block; the convolution kernel size of each convolution layer is different; The feature-guided fusion module includes a plurality of fusion enhancement branches corresponding to the number of private features; Each private feature is input into the corresponding fusion enhancement branch; Shared spatial features are input into all fusion enhancement branches simultaneously; Multi-head cross-attention mechanism is adopted in both fusion enhancement branches; The output feature decisions of each fusion enhancement branch are fused and output to the classifier.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: The memory stores a computer program; and the processor executes the target detection method according to any one of claims 1 to 7.

10. A readable storage medium storing a computer program; characterized in that: When the computer program is executed by a processor, it is used to implement the target detection method according to any one of claims 1 to 7.

Images