A target detection double-threshold adaptive regulation method based on neural feedback state machine

Abstract

The application discloses a target detection double-threshold adaptive regulation method based on a neural feedback state machine, belongs to the technical field of brain-computer fusion, and solves the problems of low coordination efficiency and poor robustness of an existing brain-computer fusion system in a dynamic environment. The method comprises the following steps: performing target detection on an image stream task, and outputting a target detection result of each target in each frame of image; the target detection result comprises a category label of the target and a corresponding confidence score of the target; receiving an electroencephalogram signal generated by an operator in a response process of the image stream task, performing multi-dimensional feature extraction on the electroencephalogram signal; fusing the results of the multi-dimensional feature extraction to obtain a comprehensive quality index; a neural feedback state machine performs adaptive regulation on a target detection double threshold according to the relationship between the comprehensive quality index and high and low thresholds of an electroencephalogram quality state; and each target is discriminated and output according to the relationship between the target detection double threshold after adaptive regulation and the confidence score.

Description

Technical Field

[0001] This invention relates to the field of brain-computer interface technology, and in particular to a target detection dual-threshold adaptive control method based on a neural feedback state machine. Background Technology

[0002] Brain-computer interface (BCI) technology combines brain-computer interfaces with artificial intelligence to achieve deep synergy between human brain intelligence and machine intelligence, constructing hybrid augmented intelligence systems with environmental adaptability and decision-making generalization capabilities. In target detection tasks with extremely high requirements for real-time performance and accuracy, such as medical diagnosis and autonomous driving, BCI systems demonstrate significant advantages.

[0003] Current brain-computer interface (BCI) target detection systems primarily employ serial, parallel, or a combination of both architectures. In a serial architecture, the AI ​​system typically performs initial detection and screening of images, followed by the final judgment of suspicious targets by a human operator; alternatively, the operator issues commands via a brain-computer interface, which the AI ​​executes. This architecture has a fixed process and lacks real-time human-machine interaction and state adaptation. In a parallel architecture, the AI ​​and the BCI process information independently and output the final result through weighted fusion, but this method struggles to dynamically adjust based on task progress and the user's brain state. While the combined serial-parallel architecture attempts to integrate the advantages of both, it still doesn't fundamentally solve the problem of insufficient system adaptability.

[0004] The existing technology has the following main drawbacks:

[0005] 1) Static task allocation and parameter settings: Key parameters such as confidence threshold and learning rate of the artificial intelligence module in the system are usually fixed values ​​that are preset and cannot be dynamically adjusted according to real-time task load, environmental changes and the operator's mental state.

[0006] 2) One-way open-loop information flow: Existing systems mostly present one-way instruction transmission from "brain to machine" or "machine to brain", lacking a central control mechanism that can evaluate the status of the "human and machine" dual system in real time and perform closed-loop optimization accordingly.

[0007] 3) Limited system synergy: Due to fixed parameters and lack of feedback, the system cannot fully utilize the automated processing capabilities of artificial intelligence when the operator is focused and the quality of the EEG signal is high. It also cannot ensure the robustness of the system by raising the machine decision threshold when the operator is fatigued or there is a lot of environmental interference. As a result, the overall performance cannot achieve the synergistic amplification of "1+1>2".

[0008] In recent years, the inference interfaces of advanced object detection models such as YOLOv6 have supported dynamic adjustment of core parameters such as confidence thresholds, providing a technical foundation for building adaptive brain-computer interface systems. Meanwhile, the time-domain and frequency-domain characteristics of event-related potentials (such as P300 components) in EEG signals have been proven to be highly correlated with an individual's attention level, cognitive load, and decision confidence, serving as reliable biological neural feedback indicators reflecting the current collaborative state of the system.

[0009] Therefore, there is an urgent need for a method that can utilize quantifiable EEG signal features as real-time neural feedback, and dynamically and adaptively regulate key parameters of artificial intelligence systems through a specific closed-loop intelligent interaction module—a neural feedback state machine—in order to build a truly intelligent, robust, and efficient brain-computer collaborative perception and decision-making system. Summary of the Invention

[0010] Based on the above analysis, the embodiments of the present invention aim to provide a target detection dual-threshold adaptive control method based on a neural feedback state machine, in order to solve the problems of low collaborative efficiency and poor robustness of existing brain-computer fusion systems in dynamic environments.

[0011] This invention provides a target detection dual-threshold adaptive control method based on a neural feedback state machine, the method comprising:

[0012] The image stream task performs object detection and outputs the object detection result for each object in each frame of the image; the object detection result includes the object's category label and its corresponding confidence score;

[0013] The system receives EEG signals generated by the operator during the response to an image stream task, performs multidimensional feature extraction on the EEG signals, and fuses the results of the multidimensional feature extraction to obtain a comprehensive quality index.

[0014] The neurofeedback state machine adaptively adjusts the dual thresholds for target detection based on the relationship between the comprehensive quality index and the high and low thresholds of the EEG quality state.

[0015] Based on the relationship between the adaptively adjusted target detection thresholds and confidence scores, each target is judged and output separately.

[0016] Based on the above solution, the present invention also makes the following improvements:

[0017] Furthermore, the adaptive adjustment of the dual thresholds for target detection is performed as follows:

[0018] make Indicates the overall quality index. , These represent the high and low thresholds of EEG quality status, respectively.

[0019] when At the same time, lower the upper limit of the target detection threshold and simultaneously raise the lower limit of the target detection threshold;

[0020] when At the same time, the upper limit of the target detection threshold is increased, and the lower limit of the target detection threshold is decreased simultaneously;

[0021] when At that time, the current target detection thresholds remain unchanged.

[0022] Furthermore, when At the same time, lower the upper limit of the target detection threshold and simultaneously raise the lower limit of the target detection threshold, and execute:

[0023] Upper limit of target detection threshold after adaptive adjustment Represented as:

[0024] (1)

[0025] in, This represents the upper limit of the current target detection threshold. The first adjustment step size for the target detection threshold;

[0026] Lower limit of target detection threshold after adaptive adjustment Represented as:

[0027] (2)

[0028] in, This is the lower limit of the current target detection threshold. This is the second adjustment step size for the target detection threshold.

[0029] Furthermore, when At the same time, increase the upper limit of the target detection threshold and simultaneously decrease the lower limit of the target detection threshold, and execute:

[0030] Upper limit of target detection threshold after adaptive adjustment Represented as:

[0031] (3)

[0032] in, This is the third adjustment step size for the target detection threshold;

[0033] Lower limit of target detection threshold after adaptive adjustment Represented as:

[0034] (4)

[0035] in, This is the fourth adjustment step size for the target detection threshold.

[0036] Furthermore, each target is evaluated and output separately, and the following steps are executed:

[0037] If confidence score C satisfy C ≥ If the current target is determined to be a confirmed target, the target detection result for the current target is output.

[0038] If confidence score C satisfy C ≤ If the current target is determined to be a non-target, it is filtered directly.

[0039] If confidence score C satisfy < C < The current target is determined to be a difficult target.

[0040] Furthermore, the receiving operator performs multidimensional feature extraction on the EEG signals generated during the response to the image stream task, and executes the following:

[0041] The system receives EEG signals generated by the operator during the response to the image stream task, preprocesses the received EEG signals, and obtains preprocessed EEG signals.

[0042] Multidimensional feature extraction was performed on the preprocessed EEG signals.

[0043] Furthermore, the multidimensional feature extraction includes amplitude, signal-to-noise ratio, and information entropy.

[0044] Furthermore, the results of the multi-dimensional feature extraction are used to obtain a comprehensive quality index, which is then executed as follows:

[0045] The extracted amplitude, signal-to-noise ratio, and information entropy are normalized respectively. The normalized amplitude, signal-to-noise ratio, and information entropy are then weighted and fused to obtain the comprehensive quality index.

[0046] Furthermore, the comprehensive quality index Represented as:

[0047] (5)

[0048] in, , , These represent the weighting coefficients for amplitude, signal-to-noise ratio, and information entropy, respectively. , , These represent the amplitude, signal-to-noise ratio, and information entropy after normalization, respectively.

[0049] further, .

[0050] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0051] (1) Enhance the collaborative efficiency of brain-computer interface systems

[0052] Dynamic adaptation to operator's neural state: By receiving EEG signals generated by the operator during the response to the image stream task and extracting multidimensional features, a comprehensive quality index is obtained. Based on this index, the dual thresholds for target detection are adaptively adjusted. This dynamic adjustment mechanism based on the operator's real-time neural state enables the brain-computer interface system to better match the operator's neural state, thereby achieving more efficient collaborative work in dynamic environments.

[0053] Enhancing the system's adaptability to complex tasks: When faced with complex image stream tasks, the operator's neural state may change. This method can adjust the target detection threshold in a timely manner according to these changes, ensuring that the system maintains good detection performance at different task stages and improving the overall adaptability of the system to complex tasks.

[0054] (2) Improve the accuracy and reliability of target detection

[0055] Optimize target detection threshold settings: Based on the relationship between the comprehensive quality index and the high and low thresholds of EEG quality status, adaptive adjustment of the dual thresholds for target detection allows for more precise determination of the target detection threshold range. When the operator's neural state is good, the upper limit of the target detection threshold is lowered while the lower limit is simultaneously raised, enabling the system to detect more potential targets and reduce missed detections. When the operator's neural state is poor, the upper limit of the target detection threshold is raised while the lower limit is simultaneously lowered, improving the reliability of the detection results and reducing false positives.

[0056] Effective target type differentiation: By setting different discrimination conditions, targets are divided into sure targets, sure non-targets, and difficult targets. For sure targets, the detection result is directly output; for sure non-targets, the detection is directly filtered; and for difficult targets, further analysis or other measures can be taken. This classification method can more accurately identify targets and improve the overall accuracy and reliability of target detection.

[0057] (3) Enhance the robustness of the system

[0058] Reduced External Interference: Because this method can dynamically adjust the target detection threshold based on the operator's neural state, the system can maintain stable operation better when faced with changes in the external environment or interference. Even when the operator's neural state fluctuates due to certain interference, the system can adjust the threshold in a timely manner through an adaptive control mechanism, reducing the impact of interference on the target detection results. This enhances the system's robustness, achieves optimal matching between the processing capabilities of both the human and machine and the task requirements, and ultimately improves the adaptive capability, task performance, and operational robustness of the entire brain-computer interface system.

[0059] Improving the system's ability to cope with abnormal situations: When the operator's mental state is abnormal, such as fatigue or lack of concentration, this method can reduce the system's sensitivity to these abnormal states by adjusting the target detection threshold, avoiding a significant drop in system performance due to the operator's poor condition, improving the system's ability to cope with abnormal situations, and further enhancing the system's robustness.

[0060] (4) Improve user experience and operational efficiency

[0061] Reduce operator workload: This method reduces the frequency of operators needing to manually adjust system parameters by adaptively adjusting the target detection threshold, thereby reducing the operator's workload and allowing the operator to focus more on the image stream task itself, thus improving operational efficiency.

[0062] Enhance operator trust in the system: Because the system can automatically adjust the target detection threshold based on the operator's neural state and can identify targets more accurately, operators will have greater trust in the system's performance and reliability, and will be more willing to use the system, thus improving the user experience.

[0063] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0064] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0065] Figure 1 This is a flowchart of the target detection dual-threshold adaptive control method based on a neural feedback state machine provided in Embodiment 1 of the present invention;

[0066] Figure 2This is a schematic diagram of the target detection dual-threshold adaptive control system based on a neural feedback state machine provided in Embodiment 2 of the present invention. Detailed Implementation

[0067] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0068] Specific embodiment 1 of the present invention discloses a target detection dual-threshold adaptive control method based on a neural feedback state machine, the flowchart of which is shown below. Figure 1 As shown, the specific explanation is as follows.

[0069] Step S1: Perform object detection on the image stream task and output the object detection result for each object in each frame of the image; the object detection result includes the object's category label and its corresponding confidence score;

[0070] Step S2: Receive the EEG signals generated by the operator during the response image stream task, and perform multi-dimensional feature extraction on the EEG signals; fuse the results of the multi-dimensional feature extraction to obtain a comprehensive quality index;

[0071] Step S3: The neurofeedback state machine adaptively adjusts the dual thresholds for target detection based on the relationship between the comprehensive quality index and the high and low thresholds of the EEG quality state.

[0072] Step S4: Based on the relationship between the adaptively adjusted target detection dual thresholds and the confidence score, each target is discriminated and output.

[0073] In the specific implementation process, steps S1-S4 can be executed sequentially; or, steps S1 and S2 can be executed simultaneously, and then steps S3 and S4 can be executed sequentially based on the results of steps S1 and S2.

[0074] The specific implementation process of each step is explained in detail below.

[0075] In step S1, object detection can be performed on the image stream task using an object detection model, outputting the object detection result for each object in each frame of the image. In specific implementation, in addition to the object's category label and its corresponding confidence score, the object detection result can also include the object bounding box position in the object detection result, depending on the actual situation.

[0076] For example, in object detection models such as YOLO, the confidence score (0≤) ≤1) It consists of the product of two key components: the probability of the target's existence and the conditional probability of the category. The specific formula is as follows:

[0077] (1)

[0078] in, The ObjectnessScore represents the probability of an object being contained within the bounding box, and its value ranges from [0,1]. This value is directly output by the model's network and normalized using the Sigmoid function. This represents the conditional probability that the target belongs to a specific category given the existence of the target, and its value ranges from [0,1]. This value is obtained by using the Softmax function on the class logits output by the model.

[0079] Preferably, step S2 can be implemented in the following manner.

[0080] Step S21: Receive the EEG signals generated by the operator during the response to the image stream task, preprocess the received EEG signals, and obtain the preprocessed EEG signals.

[0081] In practice, a non-invasive electrode cap can be used to collect the operator's EEG signals, and the collected EEG signals can be preprocessed to obtain preprocessed EEG signals.

[0082] Preferably, in this embodiment, the collected EEG signals are preprocessed, which may include operations such as amplification, filtering (e.g., 0.1-30Hz bandpass filtering) and removal of power frequency interference, thereby purifying the EEG signals and laying a good foundation for subsequent feature extraction.

[0083] Step S22: Perform multidimensional feature extraction on the preprocessed EEG signal; the multidimensional feature extraction includes amplitude, signal-to-noise ratio and information entropy.

[0084] In the specific implementation process, a specific time window (such as 300-500 ms) after the stimulus (such as the appearance of a target in the image stream) in the image stream task is used as the analysis time window. The P300 component of the EEG signal within the analysis time window is locked to carry out feature extraction.

[0085] In the specific implementation process, amplitude, signal-to-noise ratio, and information entropy can be extracted in the following ways.

[0086] Amplitude is used to calculate the absolute value of the voltage difference between the peak value of the EEG signal waveform and the average voltage of the EEG signal during the pre-stimulation baseline (200 ms before stimulation) within the analysis time window. Represented as:

[0087] (2)

[0088] in, To analyze EEG signals within a time window Voltage value at time, The average voltage of the EEG signal during the pre-stimulation baseline period (200 ms before stimulation) is given. This refers to the time range of the analysis time window for P300.

[0089] Signal-to-noise ratio (SNR) is measured by the ratio (expressed in decibels) of the signal power within an analysis time window to the noise power during the baseline period. Represented as:

[0090] (3)

[0091] in, To analyze the average power of EEG signals within a time window, The average power of the EEG signal during the pre-baseline period before stimulation.

[0092] Information entropy quantifies the complexity and regularity of EEG signals using approximate entropy or sample entropy. A higher entropy value often indicates more active brain processing and a larger information capacity. In this example, sample entropy is used, calculated based on the probability of new patterns arising in the signal sequence. The calculation formula can be simplified as follows:

[0093] (4)

[0094] Among them, sample entropy Due to the length of the subsequence Similarity tolerance EEG segment length Joint decision; Indicates length is The average proportion of brain electron sequences that meet similarity conditions; Indicates length is The average proportion of brain electron sequences that meet similar conditions.

[0095] Length is The set of subsequences:

[0096] (5)

[0097] Length is The set of subsequences:

[0098] (6)

[0099] Specifically, the subsequence length Similarity tolerance EEG segment length The relevant definitions are shown in Table 1.

[0100] Table 1. Parameters and related definitions related to sample entropy.

[0101]

[0102] Step S23: Normalize the extracted amplitude, signal-to-noise ratio and information entropy respectively, and then perform weighted fusion on the normalized amplitude, signal-to-noise ratio and information entropy to obtain the comprehensive quality index.

[0103] In the specific implementation process, in order to eliminate the dimensional differences between different features and achieve unbiased fusion of multidimensional features, the amplitude is respectively... Signal-to-noise ratio and information entropy Perform normalization, and then normalize the amplitude. Signal-to-noise ratio and information entropy The features are fused into a single comprehensive EEG feature vector. This EEG feature vector fully describes the quantitative characteristics of the current EEG signal in three orthogonal dimensions: intensity (amplitude), clarity (signal-to-noise ratio), and complexity (entropy value), and can characterize the quality of the current EEG signal and the underlying cognitive state.

[0104] Overall Quality Index It is a scalar fusion of multidimensional features, providing a single, intuitive indicator for assessing neural states and a comprehensive quality index. Represented as:

[0105] (7)

[0106] in, For the preset weight vector, , , Let represent the weighting coefficients for amplitude, signal-to-noise ratio, and information entropy, respectively, and satisfy . In practice, the weight vector can be adjusted to prioritize signal stability, clarity, and complexity based on different task scenarios. For example, for high-precision detection tasks, emphasis is placed on signal strength and clarity, while the complexity weight is reduced. For long-term monitoring tasks (such as security patrols), balance all characteristics, paying particular attention to the fatigue state reflected by the entropy value. For fast-response tasks (such as mind-controlled interaction), the emphasis is on the amplitude reflecting the intensity of attention. .

[0107] In step S3, the neural feedback state machine adaptively adjusts the target detection dual thresholds based on the relationship between the comprehensive quality index and the high and low thresholds of the EEG quality state through a preset mapping strategy, thereby achieving dynamic matching based on the operator's neural state and task requirements. The specific implementation process is described below.

[0108] The high and low thresholds of EEG quality status are respectively used as , This indicates that these two parameters can be fine-tuned depending on the specific circumstances. For example, in a certain implementation process, the preset... , .

[0109] 1) Loosening strategy under high-quality conditions

[0110] When the comprehensive quality index When, it indicates A high-quality EEG signal (large amplitude, high signal-to-noise ratio, moderate or low entropy) indicates that the image features are significant, the operator's interpretation of the target-related brain regions is consistent, and the operator is in a reliable state of high attention and low fatigue. At this point, adaptive adjustment is performed on the target detection thresholds (including the upper and lower thresholds) to perform the following operations.

[0111] Lowering the upper limit of the target detection threshold: This lowers the threshold for the machine to output a "sure target," allowing artificial intelligence to perform more screening work and reducing false negatives (missed reports) caused by excessively high thresholds. At this point, the upper limit of the target detection threshold is adaptively adjusted. Represented as:

[0112] (8)

[0113] in, This represents the upper limit of the current target detection threshold. This is the first adjustment step size for the target detection threshold, with a default value of 0.05. This metric can be fine-tuned according to the task.

[0114] Simultaneously increase the lower limit of the target detection threshold: This is used to narrow the "difficult target" range, directly classifying more low-confidence events as "certainly non-targets," thereby improving the overall system throughput efficiency. At this point, the adaptively adjusted lower limit of the target detection threshold... Represented as:

[0115] (9)

[0116] in, This is the lower limit of the current target detection threshold. This is the second adjustment step size for the target detection threshold, with a default value of 0.03. This metric can be fine-tuned according to the task.

[0117] 2) Conservative strategies under low-quality conditions

[0118] When the comprehensive quality index When, it indicates Poor EEG signal quality (small amplitude, low signal-to-noise ratio, abnormal entropy) indicates that the image features are blurry, and the operator's interpretation of brain regions related to target recognition is inconsistent, leading to operator distraction. In this case, it is necessary to investigate whether the operator is fatigued or experiencing environmental interference. At this point, adaptive adjustment of the target detection dual thresholds (including the upper and lower thresholds) is performed, and the following operations are performed.

[0119] Increase the upper limit of the target detection threshold: Raise the criterion for "confident target" and only output results with extremely high confidence to strictly suppress false detections. At this point, the upper limit of the adaptively adjusted target detection threshold... Represented as:

[0120] (10).

[0121] in, This is the third adjustment step size for the target detection threshold, with a default value of 0.05. This metric can be fine-tuned according to the task.

[0122] Simultaneously lower the lower limit of the target detection threshold: expand the range of "difficult targets," classifying more targets with moderate confidence into the category requiring further manual review or fusion judgment, relying on the system's redundancy design to ensure the reliability of the final decision. At this point, the adaptively adjusted lower limit of the target detection threshold... Represented as:

[0123] (11).

[0124] in, This is the fourth adjustment step for the target detection threshold, with a default value of 0.03. This metric can be fine-tuned according to the task.

[0125] It should be noted that, in this embodiment, the adjusted target detection dual thresholds must satisfy 0 ≤ < ≤1, 0≤ < ≤1.

[0126] 3) Stability strategy in intermediate states

[0127] When the comprehensive quality index At that time, the EEG signal is determined to be in a stable state, and the current target detection thresholds remain unchanged. That is,

[0128] (12)

[0129] (13)

[0130] In step S4, based on the target detection results of each frame of the image obtained in step S1 and the adaptively adjusted target detection dual thresholds obtained in step S3, each target can be discriminated and output according to the relationship between the adaptively adjusted target detection dual thresholds and the confidence score. That is, the target detection results of each frame of the image are adjudicated and organized. The specific implementation process is described below.

[0131] Specifically, for each objective:

[0132] 1) If satisfied C ≥ If the current target is determined to be a confirmed target, the target detection result for the current target is output.

[0133] 2) If satisfied C ≤ If the current target is determined to be a non-target, it is filtered directly.

[0134] 3) If < C < The current target is determined to be a difficult target. These targets can be marked, highlighted, or sent to a review queue to trigger secondary attentional stimulation for the operator (to regain high-quality neural state feedback), or submitted to other independent decision-making processes for final judgment.

[0135] In summary, the target detection dual-threshold adaptive control method based on neural feedback state machine proposed in this embodiment has the following significant advantages:

[0136] (1) Achieved precise control of neural feedback closed loop based on multi-dimensional physiological signals: By comprehensively calculating multiple quantifiable features such as amplitude, signal-to-noise ratio and information entropy of EEG signals, it is possible to more comprehensively and stably assess the operator's real-time cognitive state, overcome the shortcomings of single indicators being easily interfered with, and provide more accurate input basis for the dynamic adjustment of artificial intelligence parameters.

[0137] (2) A dual-threshold collaborative dynamic adjustment mechanism was constructed to optimize the division of labor between humans and machines: By simultaneously adjusting the two thresholds of "confident target" and "confident non-target", the system can more precisely control the dynamic allocation ratio of tasks among the three modes of "fully automatic processing", "human-machine collaborative processing" and "fully manual review", thus achieving the optimal balance between system efficiency and robustness.

[0138] (3) Significantly improves the overall adaptability and task performance of the brain-computer interface system: The system can automatically adapt to the dynamic changes of the operator from energetic to fatigued and the environment from good to bad, and always maintains high comprehensive performance, solving the problem of large performance fluctuations of traditional static parameter systems in changing scenarios.

[0139] (4) It provides an implementable and scalable technical framework: the proposed feature calculation method and parameter mapping strategy have clear mathematical definitions and calculation processes, which are easy to implement in engineering. This framework can be further extended to adjust other parameters such as the learning rate and number of iterations of artificial intelligence models, or applied to a wider range of brain-computer fusion scenarios such as speech recognition and brain-controlled interaction.

[0140] Specific embodiment 2 of the present invention discloses a target detection dual-threshold adaptive control system based on a neural feedback state machine. A schematic diagram of the system is shown below. Figure 2 As shown, the system includes: an EEG processing module, an artificial intelligence processing module, a neurofeedback state machine, and a fusion output module; among which,

[0141] The EEG processing module is used to receive EEG signals generated by the operator during the response to the image stream task, perform multi-dimensional feature extraction on the EEG signals, and fuse the results of multi-dimensional feature extraction to obtain a comprehensive quality index.

[0142] The artificial intelligence processing module is used to perform object detection on the image stream task and output the object detection result for each object in each frame of the image; the object detection result includes the object's category label and its corresponding confidence score;

[0143] The neurofeedback state machine is used to adaptively adjust the dual thresholds for target detection based on the relationship between the comprehensive quality index and the high and low thresholds of the EEG quality state.

[0144] The fusion output module discriminates and outputs each target based on the relationship between the adaptively adjusted target detection dual thresholds (i.e., dual threshold adjustment instructions) and the confidence score.

[0145] The system's workflow is as follows: environmental information (such as image streams) is input to the artificial intelligence processing module for preliminary analysis; simultaneously, the EEG processing module monitors and extracts the operator's EEG signals during the processing of related tasks in real time, generating a comprehensive quality index; the neurofeedback state machine generates parameter adjustment instructions (i.e., dual threshold adjustment instructions) based on the comprehensive quality index and sends them to the artificial intelligence processing module to change its processing strategy; finally, the system outputs the adaptively optimized processing results through the fusion output module.

[0146] It should be noted that the specific implementation process of this embodiment can be found in the above method embodiments, and will not be repeated here. Since this embodiment is based on the same principle as the above method embodiments, this system also has the corresponding technical effects of the above method embodiments.

[0147] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0148] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A target detection dual-threshold adaptive control method based on a neural feedback state machine, characterized in that, The method includes: The image stream task performs object detection and outputs the object detection result for each object in each frame of the image; the object detection result includes the object's category label and its corresponding confidence score; The system receives EEG signals generated by the operator during the response to an image stream task, performs multidimensional feature extraction on the EEG signals, and fuses the results of the multidimensional feature extraction to obtain a comprehensive quality index. The multidimensional feature extraction includes amplitude, signal-to-noise ratio (SNR), and information entropy; signal-to-noise ratio... Represented as: (1) in, To analyze the average power of EEG signals within a time window, The average power of the EEG signal during the pre-baseline period of stimulation; The neurofeedback state machine adaptively adjusts the dual thresholds for target detection based on the relationship between the comprehensive quality index and the high and low thresholds of the EEG quality state. The adaptive adjustment of the dual thresholds for target detection is performed as follows: make Indicates the overall quality index. , These represent the high and low thresholds of EEG quality status, respectively. when At the same time, lower the upper limit of the target detection threshold and simultaneously raise the lower limit of the target detection threshold; when At the same time, the upper limit of the target detection threshold is increased, and the lower limit of the target detection threshold is decreased simultaneously; when At the same time, the current target detection thresholds remain unchanged; Based on the relationship between the adaptively adjusted target detection thresholds and the confidence score, each target is judged and output separately; Analyze and output the results for each target separately, and then execute: If confidence score C satisfy C ≥ If the current target is determined to be a confirmed target, the target detection result for the current target is output. If confidence score C satisfy C ≤ If the current target is determined to be a non-target, it is filtered directly. If confidence score C satisfy < C < The current target is determined to be a difficult target; in, , These represent the lower limit and upper limit of the target detection threshold after adaptive adjustment, respectively.

2. The target detection dual-threshold adaptive control method based on neural feedback state machine according to claim 1, characterized in that, when At the same time, lower the upper limit of the target detection threshold and simultaneously raise the lower limit of the target detection threshold, and execute: Upper limit of target detection threshold after adaptive adjustment Represented as: (2) in, This represents the upper limit of the current target detection threshold. The first adjustment step size for the target detection threshold; Lower limit of target detection threshold after adaptive adjustment Represented as: (3) in, This is the lower limit of the current target detection threshold. This is the second adjustment step size for the target detection threshold.

3. The target detection dual-threshold adaptive control method based on neural feedback state machine according to claim 2, characterized in that, when At the same time, increase the upper limit of the target detection threshold and simultaneously decrease the lower limit of the target detection threshold, and execute: Upper limit of target detection threshold after adaptive adjustment Represented as: (4) in, This is the third adjustment step size for the target detection threshold; Lower limit of target detection threshold after adaptive adjustment Represented as: (5) in, This is the fourth adjustment step size for the target detection threshold.

4. The target detection dual-threshold adaptive control method based on neural feedback state machine according to any one of claims 1-3, characterized in that, The receiving operator performs multidimensional feature extraction on the EEG signals generated during the response to the image stream task, and executes the following: The system receives EEG signals generated by the operator during the response to the image stream task, preprocesses the received EEG signals, and obtains preprocessed EEG signals. Multidimensional feature extraction was performed on the preprocessed EEG signals.

5. The target detection dual-threshold adaptive control method based on neural feedback state machine according to claim 4, characterized in that, The results of the multi-dimensional feature extraction are used to obtain a comprehensive quality index, which is then executed as follows: The extracted amplitude, signal-to-noise ratio, and information entropy are normalized respectively. The normalized amplitude, signal-to-noise ratio, and information entropy are then weighted and fused to obtain the comprehensive quality index.

6. The target detection dual-threshold adaptive control method based on neural feedback state machine according to claim 5, characterized in that, Overall Quality Index Represented as: (6) in, , , These represent the weighting coefficients for amplitude, signal-to-noise ratio, and information entropy, respectively. , , These represent the amplitude, signal-to-noise ratio, and information entropy after normalization, respectively.

7. The target detection dual-threshold adaptive control method based on neural feedback state machine according to claim 6, characterized in that, 。

Images