Method and system for detecting existence of camouflage target based on electroencephalogram feature guidance

Abstract

The invention discloses a camouflage target existence detection method and system based on electroencephalogram feature guidance, and the method comprises the steps: constructing a camouflage target existence detection model; performing camouflage target existence judgment on the image data and the electroencephalogram data through an image branch and an electroencephalogram branch in the camouflage target existence detection model; an image encoder and an electroencephalogram encoder in the camouflage target existence detection model are used for respectively extracting image and electroencephalogram feature codes, cross-modal information alignment is realized through a decoupling fusion module, and information integrity after decoupling is ensured through a plurality of decoupling type image feature integrators and decoupling type electroencephalogram feature integrators; finally, in the reasoning stage, the existence judgment of the camouflage target is independently completed only through the image branches in the camouflage target existence detection model, and the judgment capability and generalization performance of the camouflage target existence detection model in a complex camouflage scene are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of brain-computer interface, and particularly relates to a camouflage target existence detection method and system based on electroencephalogram feature guidance. BACKGROUND

[0002] Camouflage target detection is a visual perception task with extremely high identification difficulty. The detection target is usually highly similar to the background in appearance features, presenting low contrast in color, texture, shape, etc. The blurred boundary between the foreground and the background significantly weakens the distinguishing ability of traditional computer vision models for the target, severely restricting the identification performance of the camouflage target.

[0003] The existing camouflage target detection (COD) methods are mostly based on deep convolutional neural networks, using techniques such as saliency modeling, edge feature extraction, and multi-scale fusion to enhance the perception ability of the model for low-contrast and highly concealed targets. These methods improve the detection effect of camouflage targets to a certain extent and promote the development of the field by enhancing the distinguishability of the target and the background.

[0004] At the same time, with the development of brain-computer interface and multi-modal fusion technology, existing research has begun to introduce electroencephalogram signals to assist visual task modeling and explore the cognitive features of the human brain when processing complex images. Studies have found that the human brain can quickly judge the target in low-contrast images through context integration and boundary perception, and show relevant perception responses through electroencephalogram signals. Especially in fast visual stimulation tasks, electroencephalogram signals can provide effective auxiliary criteria for deep models.

[0005] Although existing camouflage target detection methods continue to innovate in technical means, there are still many deficiencies. On the one hand, the low saliency of camouflage targets and the interference of complex environments result in weak generalization ability and robustness of the model, making it difficult to meet the actual application requirements. On the other hand, most current methods focus on target segmentation and positioning, usually assuming that the image contains camouflage targets, ignoring the judgment of target existence, affecting the computational efficiency and actual application effect of the system. In recent years, electroencephalogram signals, as an important physiological signal reflecting visual cognition of the human brain, have been introduced into the field of visual recognition, becoming a potential auxiliary information to improve the performance of camouflage target existence detection. SUMMARY

[0006] The purpose of the present application is to provide a camouflage target existence detection method and system based on electroencephalogram feature guidance.

[0007] In a first aspect, the present application provides a camouflage target existence detection method based on electroencephalogram feature guidance, which comprises: Collect electroencephalogram data corresponding to the image data; preprocess the image data and the electroencephalogram data respectively to obtain image samples and electroencephalogram samples, and construct a dataset based on the image samples and the electroencephalogram samples; Construct a camouflage target existence detection model; the camouflage target existence detection model includes an image branch and an electroencephalogram branch in parallel; the image branch includes an image encoder, an image decoupling fusion module and an image classifier connected in series; the electroencephalogram branch includes an electroencephalogram encoder, an electroencephalogram decoupling fusion module and an electroencephalogram classifier connected in series; the image encoder and the electroencephalogram encoder are respectively used to extract features of the image samples and the electroencephalogram samples to obtain high-level semantic features and spatio-temporal perception features; The image decoupling fusion module includes a plurality of decoupled image feature integrators connected in series; the plurality of decoupled image feature integrators have the same structure and each includes an image decoupling module and an image fusion module connected in series; the electroencephalogram decoupling fusion module includes a plurality of decoupled electroencephalogram feature integrators connected in series; the plurality of decoupled electroencephalogram feature integrators have the same structure and each includes an electroencephalogram decoupling module and an electroencephalogram fusion module connected in series; the number of the decoupled image feature integrators is the same as that of the decoupled electroencephalogram feature integrators; The image decoupling module includes an image private channel and an image public channel in parallel; the electroencephalogram decoupling module includes an electroencephalogram private channel and an electroencephalogram public channel in parallel; the image private channel, the image public channel, the electroencephalogram private channel and the electroencephalogram public channel are each formed by a plurality of fully connected layers and activation functions stacked in order; wherein the image public channel and the electroencephalogram public channel adopt a group of parameter-shared channel functions to process features input into the image decoupling module and the electroencephalogram decoupling module respectively to obtain image public features and electroencephalogram public features; the image private channel and the electroencephalogram private channel are two groups of independent channel functions to process features input into the image decoupling module and the electroencephalogram decoupling module respectively to obtain image private features and electroencephalogram private features; the image fusion module is used to fuse the image public features and the image private features to obtain visual domain features; the electroencephalogram fusion module is used to fuse the electroencephalogram public features and the electroencephalogram private features to obtain cognitive domain features; the image classifier and the electroencephalogram classifier are respectively used to perform target detection classification on the visual domain features and the cognitive domain features to generate prediction labels; The dataset is used to train the camouflage target existence detection model, and the image branch in the trained camouflage target existence detection model is used to discriminate the existence of camouflage targets in the image to be tested.

[0008] As preferred, the input of the image fusion module is the fusion feature obtained by the weighted fusion of the image public feature and the image private feature; in the image fusion module, the fusion feature is processed through the full connection layer and the activation function in series, and the processing result is fused with the input feature of the corresponding image decoupling module to obtain the visual domain feature output by the image fusion module; the visual domain feature is taken as the output feature of the decoupled image feature integrator; The input of the EEG fusion module is the fusion feature obtained by the weighted fusion of the EEG public feature and the EEG private feature; in the EEG fusion module, the fusion feature is processed through the full connection layer and the activation function in series, and the processing result is fused with the input feature of the corresponding image decoupling module to obtain the cognitive domain feature output by the EEG fusion module; the cognitive domain feature is taken as the output feature of the decoupled EEG feature integrator.

[0009] As preferred, in the image decoupling fusion module, the input feature of the image decoupling module in the rest of the decoupled image feature integrator except the first decoupled image feature integrator is the visual domain feature output by the last decoupled image feature integrator; the input feature of the image decoupling module in the first decoupled image feature integrator is the high-level semantic feature; the visual domain feature output by the last decoupled image feature integrator is taken as the output feature of the image decoupling fusion module. In the EEG decoupling fusion module, the input feature of the EEG decoupling module in the rest of the decoupled EEG feature integrator except the first decoupled EEG feature integrator is the cognitive domain feature output by the last decoupled EEG feature integrator; the input feature of the EEG decoupling module in the first decoupled EEG feature integrator is the spatio-temporal perception feature; the cognitive domain feature output by the last decoupled EEG feature integrator is taken as the output feature of the EEG decoupling fusion module.

[0010] As preferred, in the training process, the loss function of the camouflage target existence detection model constructed includes a private channel loss; the private channel loss is used to guide the encoder to extract the representation with domain specificity from each domain, and its expression is:

[0011] wherein, is the private channel loss; and respectively represent the image private feature and the EEG private feature extracted based on the first sample; and respectively represent the image public feature and the EEG public feature extracted based on the first sample; represents the Euclidean distance function, represents the number of samples in the training batch; This is a hyperparameter.

[0012] Preferably, the loss function of the camouflage target presence detection model further includes a common channel loss and a classification task loss; the common channel loss is used to minimize the difference in common features of samples in the visual and cognitive domains; the classification task loss evaluates the difference between the predicted label and the true label output by the camouflage target presence detection model through cross-entropy loss.

[0013] Preferably, the image encoder includes a plurality of convolutional blocks and a fully connected layer connected in sequence; each convolutional block includes a plurality of convolutional layers connected in series and a max pooling layer; the EEG encoder includes a temporal convolutional module, a depthwise convolutional module and a depthwise separable convolutional module connected in sequence.

[0014] Preferably, the EEG classifier and the image classifier have the same structure, both including multiple fully connected layers in series and activation functions.

[0015] Secondly, the present invention provides a disguised target presence detection system based on EEG feature guidance, which is used to perform the above-mentioned disguised target presence detection method; the disguised target presence detection system includes a data acquisition module, a preprocessing module, and a disguised target presence detection module; the acquisition module is used to acquire image data and corresponding EEG data; the preprocessing module is used to preprocess the acquired image data and EEG data to obtain EEG samples and image samples for constructing a dataset; the disguised target presence detection module is used to perform disguised target presence detection on the tested image.

[0016] Thirdly, the present invention provides a computer device, including 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 above-described method for detecting the presence of a camouflaged target.

[0017] Fourthly, 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 above-described method for detecting the presence of camouflaged targets.

[0018] The beneficial effects of this invention are: 1. This invention significantly improves the discriminative power and expression purity of multimodal features by extracting common and private features from EEG and image data, effectively reduces semantic interference and information redundancy between modalities, makes feature alignment between EEG and image modalities more effective, promotes the full learning and transfer of knowledge in the cognitive domain by the visual model, and thus greatly improves the accuracy and robustness of camouflaged target detection.

[0019] 2. This invention introduces EEG data into the camouflage target detection task and extracts features from image data and EEG data respectively, achieving an organic correlation between the two through a shared structure; at the same time, by introducing private channel loss and public channel loss during model training, the model gradually aligns the key discriminative information of images and EEG in semantic space during training.

[0020] 3. This invention employs a progressive fusion strategy for EEG-guided camouflage target detection. By utilizing multiple decoupled image feature integrators and dynamic weight adjustment, it fully leverages the complementary advantages of EEG and image modal features, significantly enhancing the stability and discrimination effect of multimodal fusion, and effectively improving the system's ability to identify camouflage targets and its overall detection performance in complex environments.

[0021] 4. This invention demonstrates good compatibility and stability under various mainstream image backbone networks, showing certain structural generalization ability and practical application potential. It not only provides a new multimodal fusion approach for camouflaged target detection in complex environments, but also expands the application boundaries of EEG signals in the field of target detection. Attached Figure Description

[0022] Figure 1 This is the overall flowchart of the present invention.

[0023] Figure 2 This is a schematic diagram of the electroencephalogram (EEG) acquisition process in this invention.

[0024] Figure 3 This is a schematic diagram of the camouflage target existence detection model in this invention.

[0025] Figure 4 A t-SNE visualization diagram of the output features for each semantic channel. Detailed Implementation

[0026] The present invention will be further described below with reference to the accompanying drawings.

[0027] A method for detecting the presence of camouflaged targets based on EEG features is disclosed. The camouflaged target presence detection system includes a data acquisition module, a preprocessing module, and a camouflaged target presence detection module. The acquisition module is used to acquire image data and corresponding EEG data. The preprocessing module is used to preprocess the acquired image data and EEG data to obtain EEG samples and image samples for constructing the dataset. The camouflaged target presence detection module is used to detect the presence of camouflaged targets in the tested images.

[0028] like Figure 1 As shown, the method for detecting the presence of a camouflaged target includes the following steps: Step 1: Collect multimodal data The Rapid Continuous Visual Presentation (RSVP) paradigm was used to present subjects with a sequence of camouflaged target images, sourced from the publicly available camouflaged image dataset CAMO. Figure 2 As shown, the process of acquiring raw EEG data consisted of five blocks, each containing 11 trials. Before the acquisition began, subjects were shown a set of examples containing camouflaged targets and background images to illustrate the types of targets to be identified. At the start of each trial, a gaze cross was first presented in the center of the screen, followed by a series of stimulus images presented at a frequency of 1 Hz, each image displayed for 1 second. At least three non-target images were inserted between any two target images to avoid the influence of consecutive targets. Sufficient rest periods were arranged between each block to reduce fatigue. EEG signals were recorded synchronously during the acquisition process using a NeuroScan SynAmps2 system equipped with 64-channel electrode caps and a sampling rate set to 1000 Hz. Image presentation and EEG recording were synchronized through event tagging to ensure precise timing matching.

[0029] Step 2: Data Preprocessing Raw EEG data were bandpass filtered from 0.1Hz to 40Hz, then resampled to 250Hz. A 1-second segment of the event was extracted from each image's presentation time to construct the corresponding EEG sample. Image data were uniformly resized to 224×224 pixels and normalized to ensure consistent structure and scale across input data modalities, facilitating subsequent feature extraction and coupled training. A dataset was then constructed based on the preprocessed image and EEG samples.

[0030] Step 3: Construct a model for detecting the presence of camouflaged targets like Figure 3 As shown, the camouflaged target presence detection model includes parallel image and EEG branches. The image branch includes a cascaded image encoder, an image decoupling and fusion module, and an image classifier; the EEG branch includes a cascaded EEG encoder, an EEG decoupling and fusion module, and an EEG classifier.

[0031] The image encoder consists of five convolutional blocks and two fully connected layers connected in sequence. The first two convolutional blocks have the same structure, each consisting of two convolutional layers connected in series and a max pooling layer. The last three convolutional blocks have the same structure, each consisting of three convolutional layers connected in series and a max pooling layer. The image encoder processes image samples and outputs high-level semantic features. The EEG encoder processes EEG samples through a temporal convolutional module, a depthwise convolutional module, and a depthwise separable convolutional module connected in sequence and outputs spatiotemporal awareness features.

[0032] The image decoupling and fusion module comprises five decoupled image feature integrators connected in series. These five integrators have identical structures, each including a series-connected image decoupling module and an image fusion module. Similarly, the EEG decoupling and fusion module comprises five decoupled EEG feature integrators connected in series. These integrators also have identical structures, each including a series-connected EEG decoupling module and an EEG fusion module. The image decoupling and EEG decoupling modules are used to achieve effective alignment and structured separation between image modalities and EEG modalities. The image fusion and EEG fusion modules are used to fuse features from different channels within the image decoupling and EEG decoupling modules. The image decoupling module includes parallel image private channels and image common channels; the EEG decoupling module includes parallel EEG private channels and EEG common channels. The image private channels, image common channels, EEG private channels, and EEG common channels are all constructed by sequentially stacking multiple fully connected layers and activation functions, possessing nonlinear mapping capabilities. The image common channels and EEG common channels employ a set of parameter-shared channel functions, respectively used to process the features of the input image decoupling module and the EEG decoupling module, obtaining image common features and EEG common features. This ensures that both are mapped to the same semantic space under a unified transformation structure, thereby extracting consistent cross-modal shared features and achieving consistent alignment of feature projections. The image private channels and EEG private channels are two sets of independent channel functions, respectively used to process the features of the input image decoupling module and the EEG decoupling module, obtaining image private features and EEG private features. This effectively preserves the differential information of each modality, enhances the overall recognizability and integrity of the expression, and avoids information confusion between modalities.

[0033] The image fusion module takes as input a fusion feature, which is a weighted fusion of common and private features of the image in a ratio of 0.9:0.1. Within the image fusion module, the fusion feature is processed by sequentially connected fully connected layers and activation functions. The processed result is then fused with the input features of the corresponding image decoupling module to obtain the visual domain features output by the image fusion module. These visual domain features are used as the output features of the decoupled image feature integrator. Except for the first decoupled image feature integrator, the input features of the image decoupling module in the remaining decoupled image feature integrators are the visual domain features output by the previous decoupled image feature integrator; the input features of the image decoupling module in the first decoupled feature integrator are high-level semantic features. The visual domain features output by the last decoupled image feature integrator are used as the output features of the image decoupling fusion module.

[0034] The input to the EEG fusion module is a fusion feature obtained by weighting common EEG features and private EEG features in a ratio of 0.9:0.1. Within the EEG fusion module, the fusion feature is processed through sequentially connected fully connected layers and activation functions. The processing result is then fused with the input features of the corresponding EEG decoupling module to obtain the cognitive domain features output by the EEG fusion module. These cognitive domain features are used as the output features of the decoupled EEG feature integrator. Except for the first decoupled EEG feature integrator, the input features of the decoupled EEG feature integrators in the remaining decoupled EEG feature integrators are the cognitive domain features output by the previous decoupled EEG feature integrator; the input features of the decoupled EEG feature integrator in the first decoupled EEG feature integrator are spatiotemporal perception features. The cognitive domain features output by the last decoupled EEG feature integrator are used as the output features of the decoupled EEG fusion module.

[0035] The image classifier performs object detection and classification on the visual domain features output by the image decoupling and fusion module, generating image prediction labels. The EEG classifier performs object detection and classification on the cognitive domain features output by the EEG decoupling and fusion module, generating EEG prediction labels. Both the image and EEG classifiers have the same structure, consisting of multiple fully connected layers in series, a ReLU activation function, and a softmax activation function.

[0036] Step 4: Training the model A model for detecting the presence of camouflaged targets is trained using a dataset. During training, a loss function for the model is constructed. as follows:

[0037] in, These are the weighting coefficients for each loss term, used to balance the degree of optimization between different objectives; Loss of public access routes; Losses due to private channels; Losses due to categorization of tasks.

[0038] Loss of public passageways The aim is to align common representations between the visual and cognitive domains within a shared subspace, thereby minimizing the heterogeneity gap between domains and enabling the model to extract features with cross-modal semantic consistency. Common channel loss. Represented as:

[0039] in, and They respectively represent based on the first Common image features and common EEG features extracted from each sample; Represents the Euclidean distance function; This indicates the number of samples in the training batch.

[0040] Private channel loss The aim is to guide the encoder to extract domain-specific representations from each domain, thereby improving the quality of information decoupling. Its expression is:

[0041] in, and They respectively represent based on the first Image-private features and EEG-private features extracted from each sample; It is preset margin Hyperparameters are used to control the separation distance between features.

[0042] To optimize the final recognition performance, this example introduces the standard cross-entropy loss as a task-oriented supervision signal. (Classification task loss) The difference between the model output and the true labels is evaluated and used to guide the training of the entire network, expressed as:

[0043] in, The actual label; and These are the image prediction labels and EEG prediction labels corresponding to the image samples and EEG samples, respectively.

[0044] Through this joint training strategy, this example effectively integrates knowledge from the cognitive domain into the visual domain, enabling accurate detection of camouflaged targets during the reasoning stage without relying on EEG. During training, EEG data from each subject are used sequentially to train the model. By inputting data from each subject in turn, the model is continuously updated and optimized, achieving effective learning of diverse EEG features and improving the overall expressive power and discriminative performance of cognitive features.

[0045] Step 5: Use the image branch in the trained camouflage target existence detection model to perform camouflage target existence detection on the test image.

[0046] Step Six: Method Evaluation 6-1. Evaluation Indicators Because the experimental dataset used in this example is an imbalanced dataset (fewer positive samples than negative samples), Balanced Accuracy (BA) and F1-score were employed.

[0047]

[0048] in, For accuracy; Recall rate; To balance accuracy; TP To correctly predict it as a positive class; ‌TN The correct prediction is a negative class; ‌FP The error was incorrectly predicted as a positive class. ‌FN The incorrect prediction was a negative class.

[0049] 6-2. EEG Data Assessment This example collected EEG data from 5 subjects for model training. EEGNet was used as the encoder, and a classifier with the same architecture as the invented method was connected to evaluate the EEG quality. The performance of each subject's EEG and the average performance are shown in Table 1: Table 1. EEG performance of each subject

[0050] 6-3. Performance Comparison The proposed method was systematically evaluated on several mainstream vision models, covering two categories: models specifically designed for camouflaged object detection (COD) tasks, such as SINet-V2, PFNet, and BiRefNet; and general models for image classification tasks, such as ResNet-18, ResNeXt-50, ViT-B, Swin Transformer, EfficientNet-B0, VGG16, and DenseNet-121. Considering the limited sample size of the CAMO dataset, all models were initialized using publicly available pre-trained weights to improve performance and mitigate overfitting. To ensure fairness in the evaluation, all models uniformly used the same fully connected layers as the proposed method as classifiers to classify objects based on their extracted features. Table 2 shows the Balanced Accuracy and F1-Score scores of various methods on this task.

[0051] Table 2 Performance Comparison

[0052] As shown in Table 2, VGG16 performed worse than PFNet and Swin Transformer in this task. Our invention achieved the best results across all evaluation metrics. This comparison clearly demonstrates the effectiveness of incorporating EEG cognitive information in improving the performance of camouflaged target detection tasks.

[0053] 6-4. Decoupling Strategy Analysis To verify the effectiveness of the proposed decoupling strategy, this example independently evaluated the performance of the shared and private channels of the image modality. The output features of each channel were predicted using the same fully connected classifier, and the classification results are shown in Table 3. Experimental results show that the overall performance of the fused features is better than that of the individual channel features, with the shared channel features outperforming the private channel features. This indicates that the shared channel effectively captures more discriminative cross-modal shared semantics, while the private channel retains modality-specific information. Although its performance is slightly lower, it has supplementary value in improving the overall expressive power.

[0054] Table 3 Performance comparison of each channel feature in classification tasks

[0055] Furthermore, to further verify the decoupling and sharing of feature distributions among semantic channels, this example uses t-SNE visualization to visualize the output features of each channel. The results are as follows: Figure 4 As shown, the shared channel features of the two modalities highly overlap in the projection space, indicating that the model successfully captures modality-independent shared representations in the shared channels; while the private channel features remain clearly separated, reflecting the unique semantic structure of each modality. This structural validation supports the effectiveness and rationality of the proposed method.

[0056] 6-5. Generalization Analysis To verify the generality of the proposed multimodal coupling method, this example evaluates the classification performance of the multimodal model by changing the image modality feature extractor while keeping the EEG branches unchanged, and compares it with the visual results of the independent vision model. As shown in Table 4, the proposed method outperforms the corresponding single-modal model in all combinations, demonstrating good structural generalization ability and model versatility.

[0057] Table 4 Performance comparison of the present invention under different visual backbones

[0058] 6-6. Ablation Studies To verify the effectiveness of the key structures in the proposed multimodal coupling framework, two sets of ablation experiments were designed in this example to evaluate the impact of the decoupling strategy and the progressive fusion method on the overall performance. The results are shown in Table 5.

[0059] First, to analyze the effect of the decoupling strategy, a control model was constructed for this example. In this setting, the primary features of the image and EEG modalities were no longer divided into shared and private channels, but were directly aligned, while other structures remained unchanged. Experimental results show that the control model performed significantly worse than the full model on multiple metrics. This indicates that the channel decoupling mechanism plays a crucial role in separating modal-shared and modal-specific information, helping to improve the model's semantic expressiveness and cross-modal modeling accuracy.

[0060] Furthermore, to evaluate the contribution of the progressive fusion method, a model version with direct fusion without iteration was constructed. In this setting, shared and private features between images and EEG are no longer fed back to the primary features, but are directly fed into the classifier for discrimination. As shown in Table 5, the performance also shows a certain degree of decline, indicating that the progressive fusion strategy has a significant effect on mitigating potential information loss during feature decoupling and enhancing intermodal collaborative modeling.

[0061] In summary, both ablation experiments verified the effectiveness of the proposed channel decoupling and progressive fusion mechanism, which is a key structural design for improving the performance of multimodal coupled learning.

[0062] Table 5 Comparison of ablation test results for key structures

[0063] 6-7. Other preliminary experiments To determine the settings for each key hyperparameter, several pre-experiments were conducted in this sample before formal training.

[0064] (1) To further explore the impact of the number of decoupled feature integrators, this example varied the number of decoupled feature integrators in the model. As shown in Table 6, the model achieved optimal performance when the number of decoupled feature integrators was 5, i.e., PFI = 5. When PFI = 3, the performance was sometimes even worse than without any feedback, indicating that early feedback may introduce unstable gradients due to imperfect fused representations. In contrast, a moderate number of feedback iterations (e.g., PFI = 5) enabled more effective back-instruction and stable learning. Further increasing the number of iterations led to diminishing returns and may even have resulted in overfitting.

[0065] Based on these findings, this example sets PFI = 5 as the default configuration in the final model to balance training stability and feedback effectiveness.

[0066] Table 6. Effect of different iteration numbers

[0067] (2) To ensure balanced and unbiased fusion within each modality, this example sets the same public and private feature fusion weights across the EEG and visual modalities. This design avoids introducing artificial fusion bias between modalities, promotes fair comparisons, and facilitates a clearer analysis of the fusion effect within each modality.

[0068] This example performs a grid search on the candidate fusion weight set {0.1, 0.3, 0.5, 0.7, 0.9}, where each value α represents the weight assigned to the private features, while the public features are weighted by 1-α.

[0069] Table 7 Performance under different fusion weight settings

[0070] As shown in Table 7, the model achieves optimal performance when the private feature weight α is set to 0.1, indicating that emphasizing common features leads to more accurate target presence detection. As α increases, the balance between accuracy and F1 score gradually decreases, with the lowest result observed at α=0.9. This trend suggests that common features, which capture modality invariance and task-related information, are more reliable for classification, while over-reliance on private features may introduce noise or overfitting.

Claims

1. A method for detecting the presence of camouflaged targets based on EEG features, characterized in that: The method includes: Collect EEG data corresponding to image data; preprocess the image data and EEG data respectively to obtain image samples and EEG samples, and construct a dataset based on the image samples and EEG samples; A camouflage target existence detection model is constructed; the camouflage target existence detection model includes a parallel image branch and an EEG branch; the image branch includes a cascaded image encoder, an image decoupling and fusion module, and an image classifier; the EEG branch includes a cascaded EEG encoder, an EEG decoupling and fusion module, and an EEG classifier; the image encoder and the EEG encoder are used to extract features from image samples and EEG samples, respectively, to obtain high-level semantic features and spatiotemporal perception features; The image decoupling and fusion module includes multiple decoupled image feature integrators connected in series; the multiple decoupled image feature integrators have the same structure, each including a series-connected image decoupling module and an image fusion module; the EEG decoupling and fusion module includes multiple decoupled EEG feature integrators connected in series; the multiple decoupled EEG feature integrators have the same structure, each including a series-connected EEG decoupling module and an EEG fusion module; the number of decoupled image feature integrators and decoupled EEG feature integrators is the same; The image decoupling module includes parallel image private channels and image common channels; the EEG decoupling module includes parallel EEG private channels and EEG common channels; the image private channels, image common channels, EEG private channels, and EEG common channels are all composed of multiple fully connected layers and activation functions stacked sequentially; wherein, the image common channels and the EEG common channels use a set of parameter-shared channel functions, which are used to process the features of the input image decoupling module and the EEG decoupling module respectively to obtain image common features and EEG common features; the image private channels and the EEG private channels are two sets of independent channel functions, which are used to process the features of the input image decoupling module and the EEG decoupling module respectively to obtain image private features and EEG private features; the image fusion module is used to fuse the image common features and image private features to obtain visual domain features; the EEG fusion module is used to fuse the EEG common features and EEG private features to obtain cognitive domain features; the image classifier and the EEG classifier are used to perform target detection and classification on the visual domain features and cognitive domain features respectively to generate predicted labels; A model for detecting the presence of camouflaged targets is trained using a dataset, and the image branch of the trained model is used to determine the presence of camouflaged targets in the tested image.

2. The method for detecting the presence of camouflaged targets based on EEG features according to claim 1, characterized in that: The input to the image fusion module is the fusion feature obtained by weighted fusion of image public features and image private features. In the image fusion module, the fusion feature is processed by a series of fully connected layers and activation functions, and the processing result is fused with the input features of the corresponding image decoupling module to obtain the visual domain features output by the image fusion module. The visual domain features are used as the output features of the decoupled image feature integrator. The input to the EEG fusion module is the fusion feature obtained by weighted fusion of EEG public features and EEG private features. In the EEG fusion module, the fusion feature is processed by a series of fully connected layers and activation functions, and the processing result is fused with the input features of the corresponding image decoupling module to obtain the cognitive domain features output by the EEG fusion module. The cognitive domain features are used as the output features of the decoupled EEG feature integrator.

3. The method for detecting the presence of camouflaged targets based on EEG features according to claim 2, characterized in that: In the image decoupling and fusion module, except for the first decoupling image feature integrator, the input features of the image decoupling module in the other decoupling image feature integrators are the visual domain features output by the previous decoupling image feature integrator; the input features of the image decoupling module in the first decoupling feature integrator are high-level semantic features; and the visual domain features output by the last decoupling image feature integrator are used as the output features of the image decoupling and fusion module. In the EEG decoupling fusion module, except for the first decoupling EEG feature integrator, the input features of the EEG decoupling module in the other decoupling EEG feature integrators are the cognitive domain features output by the previous decoupling EEG feature integrator; the input features of the EEG decoupling module in the first decoupling EEG feature integrator are spatiotemporal perception features; and the cognitive domain features output by the last decoupling EEG feature integrator are used as the output features of the EEG decoupling fusion module.

4. The method for detecting the presence of camouflaged targets based on EEG features according to claim 1, characterized in that: During training, the loss function of the constructed camouflaged target presence detection model includes a private channel loss; the private channel loss is used to guide the encoder to extract domain-specific representations from each domain, and its expression is: in, Losses due to private channels; and They respectively represent based on the first Image-private features and EEG-private features extracted from each sample; and They respectively represent based on the first Common image features and common EEG features extracted from each sample; Represents the Euclidean distance function. Indicates the number of samples in the training batch; This is a hyperparameter.

5. The method for detecting the presence of camouflaged targets based on EEG features according to claim 4, characterized in that: The loss function of the camouflage target presence detection model also includes a common channel loss and a classification task loss; the common channel loss is used to minimize the difference in common features of samples in the visual and cognitive domains; the classification task loss evaluates the difference between the predicted label and the real label output by the camouflage target presence detection model through cross-entropy loss.

6. The method for detecting the presence of camouflaged targets based on EEG features according to claim 1, characterized in that: The image encoder includes multiple convolutional blocks and fully connected layers connected in sequence; each convolutional block includes multiple convolutional layers connected in series and a max pooling layer; the EEG encoder includes a temporal convolutional module, a depthwise convolutional module, and a depthwise separable convolutional module connected in sequence.

7. The method for detecting the presence of camouflaged targets based on EEG characteristics according to claim 1, characterized in that: The EEG classifier and the image classifier have the same structure, both including multiple fully connected layers in series and activation functions.

8. A system for detecting the presence of camouflaged targets based on EEG characteristics, characterized in that: This system is used to execute the method for detecting the presence of camouflaged targets based on EEG features as described in claim 1; the camouflaged target presence detection system includes a data acquisition module, a preprocessing module, and a camouflaged target presence detection module; The acquisition module is used to acquire image data and corresponding EEG data; the preprocessing module is used to preprocess the acquired image data and EEG data to obtain EEG samples and image samples for constructing the dataset; the camouflage target presence detection module is used to detect the presence of camouflage targets in the tested image.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: The memory stores a computer program; the processor executes the method for detecting the presence of a camouflaged target as described in any one of claims 1-7.

10. A readable storage medium storing a computer program; characterized in that: When the computer program is executed by the processor, it is used to implement the method for detecting the presence of camouflaged targets as described in any one of claims 1-7.

Images