Methods for studying the brain cognitive mechanisms of aircraft recognition in infrared background based on electroencephalography (EEG)

Abstract

This invention discloses a method for studying the brain's cognitive mechanism of aircraft recognition in an infrared background based on electroencephalography (EEG). The method includes: selecting experimental subjects and determining the experimental procedure; preprocessing the experimental data; decoding and classifying the preprocessed data using machine learning; performing source analysis on the EEG; and constructing a Granger causal network to study the collaboration between various brain regions during the aircraft recognition cognitive process. This invention focuses on the object recognition process of an aircraft (a man-made object) in an infrared background. Based on task-state EEG data from an aircraft recognition experiment, a brain network is constructed to attempt to analyze the cognitive mechanism of human brain recognition of aircraft.

Description

Technical Field

[0001] This invention relates to the fields of machine learning and EEG cognitive technology, and in particular to a method for studying the brain cognitive mechanism of aircraft recognition in infrared background based on EEG. Background Technology

[0002] In modern warfare characterized by information and technology, air superiority holds paramount importance. Precisely striking enemy aircraft during combat carries profound strategic significance, urgently requiring guided weapon systems to possess the ability to accurately identify aircraft in real time. Currently, aircraft identification technology in weapon systems primarily relies on machine vision, processing and identifying images containing aircraft against an infrared background. Infrared detection imaging technology, due to its advantages such as high sensitivity, strong resistance to electromagnetic interference, good environmental adaptability, large dynamic range, and all-weather operation, has been widely applied in military scenarios such as infrared guidance. However, military scenarios are often highly complex; targets such as aircraft typically move at high speeds and are small, making the accuracy and real-time performance of machine vision limited by image noise and numerous uncertainties such as aircraft attitude and size.

[0003] Compared to traditional machine vision, the human brain can react more quickly based on theoretical knowledge and practical experience stored within it. This is precisely the reason and value of brain-inspired research in my country's brain science program. Brain-inspired research can also be applied to the recognition of objects such as airplanes. Understanding the cognitive mechanisms of the human brain in object recognition will help deepen the understanding and optimize neural networks. Therefore, in-depth exploration of the cognitive neural mechanisms of the human brain in airplane recognition, forming brain-inspired models, and helping to optimize related neural network algorithms may effectively improve the ability of guided weapons to accurately identify airplanes. To this end, preliminary research needs to fully explore the cognitive process of the brain in airplane recognition. Only by fully understanding its cognitive neural mechanisms can a fundamental basis be provided for subsequent brain-inspired modeling.

[0004] Object recognition in the brain requires the involvement of the visual system, and early research primarily focused on this aspect. David Hubel and Torsten Wiesel initially discovered that information processing in the visual cortex, located around the calcarine fissure in the occipital lobe of the brain, is hierarchical. With the deepening of research over the following decades, the visual cortex of the brain has been divided into the primary visual cortex (V1, also known as the striate cortex) and the extrastriate cortex (V2, V3, V4, V5). V1's functions include recognizing edges, short lines, and line endpoints. Its output information has two pathways: the ventral pathway (VP) and the dorsal pathway (DP). The ventral pathway originates from V1, passes through V2 and V4, and enters the inferior temporal lobe, playing a crucial role in object recognition. The dorsal pathway primarily assists in motion recognition.

[0005] Beyond studying the fundamental physiological structures of the brain for object recognition, research has further revealed that the higher visual cortex, specifically the ventral occipital temporal cortex (VOTC), processes information about different objects differently, corresponding to the activation of different subregions within the VOTC. Some studies suggest this phenomenon is a result of bottom-up information flow, meaning that the diverse visual features input from the visual level lead to different activation patterns in different brain regions within the VOTC for different object categories. Other studies, comparing the object recognition performance of congenitally blind individuals and sighted controls, have found that this phenomenon is not entirely driven by visual features; furthermore, comparing the brain's mechanisms for recognizing animals and man-made objects reveals that the brain regions responsible for animal recognition rely more heavily on visual input, while those for man-made object recognition do not depend on visual experience.

[0006] Researchers have begun to consider that when faced with objects that do not rely on visual experience, the human brain often combines its own memory and experience, and this working mechanism may involve more brain regions. A study on the object recognition and comprehension abilities of brain-injured patients found that damage to some white matter in the frontal lobe of these patients significantly affected their object recognition and comprehension abilities, suggesting that object recognition and comprehension may depend on structural connections between distant brain regions.

[0007] In general, the brain employs different information processing mechanisms for recognizing various objects. While it is primarily based on the ventral pathway, it often requires the joint participation of higher cortical layers (such as the frontal lobe) to construct more complex networks with the primary visual cortex and the ventral occipital-temporal cortex, enabling object recognition. In-depth research into the brain's object recognition cognitive mechanisms allows researchers not only to decode and classify various objects based on cognitive activities but also to build various computer network models that mimic the brain's iterative and abstract visual information processing mechanism from lower to higher levels. This leads to the development of artificial intelligence algorithms with greater interpretability and practical value.

[0008] However, current research on object recognition and cognitive processes is mostly based on functional magnetic resonance imaging (fMRI). While fMRI provides valuable insights into brain cognitive activity and brain-like research, it lacks understanding of the evolution of cognitive activity over time in object recognition. Furthermore, it could provide more framework references for brain-like models based on large-scale functional network connectivity. Most research based on electroencephalogram (EEG) signals primarily uses the frequency and time-frequency characteristics of EEG to decode various objects, especially using event-related potential (ERP) features and inherent machine learning or deep learning algorithms for object recognition. Exploring the specific cognitive mechanisms of the brain mainly involves spectral analysis to discover the enhancement of gamma band activity during object recognition. Therefore, constructing functional connectivity networks between different brain regions based on EEG signals could be used to study the cognitive processes of object recognition. Summary of the Invention

[0009] To address the problems existing in the prior art, the purpose of this invention is to provide a method for studying the brain cognitive mechanism of aircraft recognition in an infrared background based on electroencephalography (EEG). This invention focuses on the object recognition process of an aircraft, a man-made object, in an infrared background. Based on task-state EEG during the aircraft recognition experiment, a brain network is constructed to attempt to analyze the cognitive mechanism of human brain recognition of aircraft.

[0010] To achieve the above objectives, the technical solution adopted by this invention is: a method for studying the brain cognitive mechanism of aircraft recognition in infrared background based on electroencephalography (EEG), comprising the following steps:

[0011] Step 1: Select experimental subjects and determine the experimental procedure;

[0012] Step 2: Preprocess the experimental data;

[0013] Step 3: Decode and classify the preprocessed data using machine learning;

[0014] Step 4: Perform source analysis on the EEG;

[0015] Step 5: Construct Granger causal network. By constructing the brain network, we can study the collaboration between various brain regions during the process of aircraft recognition and cognition.

[0016] As a preferred embodiment, the experimental procedure in step 1 is as follows:

[0017] Infrared background images containing aircraft targets of different sizes and attitudes, as well as environmental interference factors, were used as target images for testing, while infrared background images containing only environmental interference factors and no aircraft were used as control images. Subjects were required to press a button with their dominant hand when they identified a target image containing an aircraft, and not to press a button when they identified a non-target image. The ratio of the target image to the control image was 1:2, and each group was played repeatedly at certain time intervals.

[0018] As a preferred implementation, during the experiment, the EEG signal was acquired using an EGI 64-lead EEG acquisition device with GSN-Hydrocel64 electrode distribution, a sampling rate of 1kHz, and electrode impedance kept below 30KΩ during the acquisition process.

[0019] In a preferred embodiment, step 2 specifically includes:

[0020] First, subjects with problematic data were excluded; second, artifacts from fMRI were removed using NetStation software accompanying the EGI device; and then, EEGLAB was used for further preprocessing.

[0021] As a preferred implementation, further preprocessing using EEGLAB specifically includes the following steps:

[0022] (1) Locate the electrodes and remove four useless electrodes, including E32 and E43 used for ECG acquisition and E62 and E63 used for EEG acquisition.

[0023] (2) The faulty conductor with poor contact is replaced by interpolation using the average signal value of the three surrounding electrodes;

[0024] (3) A 0.3-45Hz FIR bandpass filter is used for filtering, with a notch filter of 50Hz;

[0025] (4) Downsampling: reduce the sampling rate to 500Hz to improve the calculation speed;

[0026] (5) Re-reference: Convert the reference method of each channel EEG to an average reference;

[0027] (6) Segmentation: Extract the data segment of correct response. Correctly identifying the airplane and pressing the key is defined as yes1; correctly identifying that there is no airplane and not pressing the key is defined as no1. The (-20, 300) ms before and after the picture stimulus in the two cases of yes1 and no1 are used as the target data segment and baseline correction is performed.

[0028] (7) Delete bad segments and identify and remove artifacts based on ICA to finally obtain the number of valid data segments.

[0029] As a preferred embodiment, step 3 is specifically as follows:

[0030] The time-domain EEG signals under the "yes1" and "no1" conditions were used as two types of samples, and samples were taken at various time points for model training to classify whether an aircraft was identified. The machine learning model used was the SVM model provided by Brainstorm, and the model was validated using three-fold cross-validation. In addition, to ensure the balance of the two types of samples, the number of data segments under both conditions was made the same through random sampling before model training. Finally, the model was evaluated based on the classification accuracy.

[0031] In a preferred embodiment, step 3 further includes:

[0032] The whole brain was divided into three parts: anterior, middle, and posterior. Multichannel EEG time-domain signals from the whole brain and the three parts were decoded to explore whether other remote brain regions, besides the occipital-temporal lobe located in the posterior part of the brain, contribute to the effective recognition of airplanes during the process of airplane recognition.

[0033] As a preferred implementation, in step 4, the Brainstorm toolbox is used to perform source tracing and inversion on each segment of data from each subject using a distributed source model, specifically including the following steps:

[0034] First, the source space is constrained to the cerebral cortex, and the head model is calculated using the OpenMEEG BEM algorithm. Then, the source is estimated using the minimum norm estimation method, and normalized using the sLORETA algorithm. At the same time, the cerebral cortex is divided into 68 brain regions using the Desikan-Killiany atlas.

[0035] As a preferred implementation, in step 5, the Brainstorm toolkit based on MATLAB is used to construct a 68*68 Granger causal network, i.e., the GC network connection matrix, by using the 68 brain regions traced back to their origins as network nodes.

[0036] The beneficial effects of this invention are:

[0037] This invention demonstrates through machine learning decoding and brain network analysis that aircraft recognition requires not only the participation of the occipital-temporal visual system, but also more distant brain regions such as the parietal lobe and cingulate gyrus. Furthermore, this invention also proves that the human brain can achieve effective aircraft recognition in a very short time, and its cognitive neural mechanism is of great significance for the subsequent construction of brain-like models to enable machines to quickly recognize aircraft. Attached Figure Description

[0038] Figure 1 This is an example image of the experimental stimulus in an embodiment of the present invention. Figure 1 (a) is a picture of the target containing an airplane, and (b) is a comparison picture without an airplane.

[0039] Figure 2 This is a schematic diagram of brain partitioning in an embodiment of the present invention;

[0040] Figure 3 This is a graph showing the machine learning classification accuracy curves for yes1 and no1 at various time points in this embodiment of the invention. Figure 3 (a) is a classification result curve based on whole-brain EEG signals, (b) is a classification result curve based on EEG signals from the anterior brain region, (c) is a classification result curve based on EEG signals from the middle brain region, and (d) is a classification result curve based on EEG signals from the posterior brain region.

[0041] Figure 4 This is a group-averaged Granger causal network connection matrix diagram for three stages (0-100ms, 100-200ms, and 200-300ms) in two cases where an aircraft is identified (yes1) or (no1). Figure 4 In the middle, (a) corresponds to yes 10-100ms, (b) corresponds to yes 1100-200ms, (c) corresponds to yes 1200-300ms, (d) corresponds to no 10-100ms, (e) corresponds to no 1100-200ms, and (f) corresponds to no 1200-300ms.

[0042] Figure 5 This is the group-averaged Granger causal network topology diagram for the three stages (0-100ms, 100-200ms, and 200-300ms) in the two cases of aircraft presence (yes1) and absence (no1) identified in this embodiment of the invention. Figure 5 In the middle, (a) corresponds to yes 10-100ms, (b) corresponds to yes 1100-200ms, (c) corresponds to yes 1200-300ms, (d) corresponds to no 10-100ms, (e) corresponds to no 1100-200ms, and (f) corresponds to no 1200-300ms. Detailed Implementation

[0043] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0044] Example

[0045] A method for studying the brain cognitive mechanisms of aircraft recognition in infrared background based on electroencephalography (EEG) includes:

[0046] 1. Experimental subjects and experimental procedures:

[0047] All subjects in this experiment were undergraduate students from the Fourth Military Medical University, and each subject underwent a basic information questionnaire. The basic information questionnaire included age, gender, dominant hand, color blindness, smoking, drinking, family history of mental illness, and history of head injury. Based on the basic information questionnaire, 31 subjects were selected who were right-handed, not colorblind, did not smoke or drink alcohol, and had no family history of mental illness or head injury.

[0048] like Figure 1 As shown, the experiment used infrared background images containing aircraft targets of different sizes and attitudes, as well as environmental interference factors, for testing. Infrared background images containing only environmental interference factors but no aircraft were used as a control. Participants were required to press a button with their dominant hand when they identified an image containing an aircraft target; otherwise, they did not press a button. A total of 30 images were used, with a target image to control image ratio of 1:2, and were played in a loop three times, with a 30-second interval between each group. During the experiment, EEG signals were acquired using an EGI 64-lead EEG acquisition device with GSN-Hydrocel64 electrodes, a sampling rate of 1kHz, and electrode impedance maintained below 30KΩ during acquisition.

[0049] 2. Data preprocessing:

[0050] First, two subjects with problematic data were excluded, and finally 29 subjects were included for analysis. In addition, since EEG was collected synchronously with fMRI, artifacts caused by fMRI were removed using the NetStation software配套软件 of the EGI device. Then, further preprocessing was performed using EEGLAB, mainly including the following steps: (1) Locating electrodes and deleting four useless electrodes (here, E32 and E43 were used to collect electrocardiogram, and E62 and E63 were used to collect electrooculogram); (2) Interpolating and replacing the bad leads with poor contact using the average signal of the surrounding three electrodes; (3) Filtering (FIR band-pass filtering of 0.3 - 45 Hz and 50 Hz notch filtering); (4) Downsampling to reduce the sampling rate to 500 Hz to improve the calculation speed; (5) Re-referencing to convert the reference mode of each channel's EEG to average reference; (6) Segmenting, intercepting the data segments of correct responses (including correctly identifying an aircraft and pressing the key, defined as yes1; correctly identifying no aircraft and not pressing the key, defined as no1), taking the (-20, 300) ms before and after the picture stimulus in both the yes1 and no1 cases as the target data segments and performing baseline correction; (7) Deleting bad segments and identifying and removing artifacts based on ICA. Finally, the number of valid data segments: there were 617 valid data segments in the yes1 case and 1310 valid data segments in the no1 case.

[0051] 3. Machine learning decoding:

[0052] To study the cognitive process of aircraft recognition in the infrared background, the time-domain EEG signals in the yes1 and no1 cases were directly used as two types of samples, and samples were taken at each time point for model training to classify whether an aircraft was recognized. The machine learning model used the SVM model provided by Brainstorm, and the model was verified through three-fold cross-validation. In addition, to ensure the balance of the two types of sample sizes, the number of data segments in both cases was made 617 through random sampling before model training. Finally, the model was evaluated based on the classification (decoding) accuracy.

[0053] At the same time, to preliminarily explore the collaborative participation mechanism of each region of the brain in the cognitive process, as Figure 2 shown, the whole brain was roughly divided into three parts: front, middle, and back. Decoding was performed based on the multi-channel EEG time-domain signals in the whole brain and within the three partitions respectively to preliminarily explore whether there are other remote brain regions that contribute to the effective recognition of aircraft in addition to the participation of the occipitotemporal lobe located in the posterior part of the brain during the aircraft recognition process of the brain.

[0054] 4. EEG source analysis:

[0055] The EEG signals collected by the scalp are actually the superposition of signals released by many neurons in the brain on the recording electrodes. To better map the brain's neural activity to the cerebral cortex, source analysis is needed to inversely deduce the location, direction, and intensity of the activity source in the cerebral cortex based on the scalp EEG data. In practical applications, source analysis of EEG is approximated as a linear problem, i.e., Y = AX, where: Y is the signal actually recorded by the scalp; X is the source information vector to be spatially located; and A is the transfer matrix (gain matrix), reflecting the transmission of signals by cerebrospinal fluid, meninges, skull, scalp, etc., which can be obtained by constructing a suitable head model.

[0056] Common methods for source tracing analysis fall into two main categories: equivalent current dipole models and distributed source models. The former is best suited for early responses within the post-stimulus latency period and cannot generalize to capture complex dynamics over long periods, while the latter is more widely applicable. This embodiment uses the Brainstorm toolbox to perform source tracing inversion on data from various subjects using a distributed source model. First, the source space is constrained to the cerebral cortex, and the OpenMEEG BEM algorithm is used to calculate the head model. Then, the minimum norm estimation method is used for source estimation, followed by normalization using the sLORETA algorithm. Simultaneously, the Desikan-Killiany atlas is used to divide the cerebral cortex into 68 brain regions. Thus, source tracing analysis inverts the EEG signals from 60 electrode channels on the head surface to 68 brain regions in the cerebral cortex, which is more conducive to subsequent analysis and interpretation after constructing a brain network connectivity matrix based on cortical brain regions. 5. Construction of Granger causal networks:

[0057] By constructing a brain network, this study explores the collaboration between various brain regions during the aircraft recognition and cognition process. Compared to undirected networks, directed causal networks can further provide information flow direction. Therefore, this embodiment utilizes the Brainstorm toolkit based on MATLAB to construct a 68*68 Granger causal network (GC network) connection matrix, using the 68 traced brain regions as network nodes.

[0058] Granger causality, proposed by Clive Granger, is a measure of linear correlation. It examines whether the variance of the linear autoregressive model estimate of signal x(t) decreases when a second signal y(t) is added to the linear model estimate. If it does, then signal y(t) has a Granger causal effect on the first signal x(t), meaning that past independent information of y(t) improves the predictability of x(t) beyond the information contained in the past of x(t). The measure of Granger causality is non-negative and zero when there is no Granger causality.

[0059] The Granger causality is calculated as follows:

[0060] First, x(t) is modeled using a linear AR model in two ways, where p is the model order and represents the amount of past information included in the prediction of future samples:

[0061]

[0062]

[0063] The calculation method for Granger causality is then obtained as follows:

[0064]

[0065] The following is a further explanation of this embodiment:

[0066] 1. Decoding the aircraft recognition and cognition process against an infrared background:

[0067] For the two scenarios of identifying whether an aircraft target exists in the infrared image (i.e., yes1 and no1), the classification results based on the time-domain EEG signal data of the whole brain, anterior, middle, and posterior brain regions at each sampling time point are as follows: Figure 3 As shown.

[0068] In classification results based on whole-brain features, the decoding accuracy shows an overall trend of gradual increase, which can be roughly divided into three stages: 0-100ms, 100-200ms, and 200-300ms.

[0069] (1) Stage 1: 0-100ms, the decoding accuracy based on the whole brain is generally below 60%, but a significant peak appears at 50ms with an accuracy of 81.5%. Around 50ms, the accuracy curves based on the three partitions also show corresponding peaks, with the decoding accuracy based on the front and back parts reaching a significant peak of 70.5%, while the accuracy based on the middle part, although having a peak, is below 60%. In addition, around 100ms, the decoding accuracy based on the whole brain and the three partitions begins to rise significantly, with the decoding accuracy based on the back part first reaching a peak of 80.5% at 98ms.

[0070] (2) Phase Two: 100-200ms. The overall decoding accuracy based on the whole brain increases from 75% at 100ms to 93% at 200ms, with peaks of 79% at 106ms and 84% at 148ms. The accuracy trend based on the front part is very similar to that based on the whole brain, increasing from 62.5% at 100ms to 86.5% at 200ms, with peaks of 70% at 104ms and a significant peak of 92% at 150ms. Although the decoding accuracy based on the back part shows an overall upward trend, the changes are not significant after fluctuations, and there is no peak at 150ms, but rather a peak of 76.5% at 138ms. The decoding accuracy based on the middle part shows an overall downward trend, except for a significant peak of 80.5% at 110ms.

[0071] (3) Stage 3: 200-300ms. The decoding accuracy based on the whole brain basically stabilizes, especially in the first 50ms, where the accuracy is generally above 90%, reaching a maximum of 97.5%. Although it declines slightly around 250ms, it eventually stabilizes at around 90%. The decoding accuracy based on the front part follows a similar trend to the whole brain, reaching a maximum of 94.5% in the first 50ms, and stabilizing at around 85% after a decline. The decoding accuracy based on the back part is less similar to the whole brain, reaching a maximum of 92% in the first 50ms, and stabilizing at around 80% after a decline. The decoding accuracy based on the middle part drops completely below 50%.

[0072] 2. Brain network analysis of aircraft recognition process under infrared background:

[0073] To further explore the collaboration among different brain regions, based on the decoding results in three phases (0-100ms, 100-200ms, and 200-300ms), a GC network connectivity matrix was constructed for each subject's data segments in both cases of recognizing an airplane (yes1) and not recognizing an airplane (no1), using source signals from 68 brain regions. The group-averaged network connectivity matrices for the three phases in both cases are shown below. Figure 4 As shown, Figure 4 In the diagram, F represents the frontal lobe, P the parietal lobe, C the central region, T the temporal lobe, O the occipital lobe, and L the cingulate gyrus. The network connection direction is from the horizontal brain regions of the matrix to the vertical brain regions of the matrix.

[0074] Comparing the connection matrices between yes1 and no1 at different time stages reveals that in the connection matrix where no airplane is identified, the stronger GC connections are mainly connections from multiple brain regions to the central region. In the connection matrix where the airplane is correctly identified, not only are the corresponding GC connections from multiple brain regions to the central region stronger, but there are also strong GC connections between other brain regions. Comparing the connection matrices of the three different stages in the yes1 and no1 cases shows that as the identification process progresses, GC connections (especially those pointing to the central region) gradually become stronger, with the most significant enhancement of the GC network in the 200-300ms stage. An exception is in the yes1 case, where, compared to 100-200ms, the connections from multiple brain regions to the frontal and parietal lobes are stronger in the 0-100ms connection matrix.

[0075] To understand the primary information flow during collaboration among different brain regions, strong GC connections (those with GC values ​​greater than 95% of the maximum GC value in the connection matrix) are further extracted from the network connectivity matrix for each case and time stage, and a brain network connectivity topology map is drawn, such as... Figure 5 As shown.

[0076] When the aircraft was not recognized, the origins of strong GC connections in all three stages were primarily located in the posterior brain (occipital lobe, parietal lobe, posterior cingulate gyrus, and temporal lobe), and the endpoints were all in the right paracentral lobule of the central region, with a connection between the left paracalcarine gyrus of the occipital lobe and the right paracentral lobule. Furthermore, the maximum GC connectivity increased with the progress of recognition (0-100ms: GCmax = 11.5; 100-200ms: GCmax = 11.6; 200-300ms: GCmax = 12.1), which may indicate a continuous enhancement of information exchange during the cognitive process.

[0077] When an airplane is detected, the maximum GC connectivity also gradually increases with the progress of detection and is greater than the corresponding value when the airplane is not detected at any stage (0-100ms: GCmax = 12.1; 100-200ms: GCmax = 12.9; 200-300ms: GCmax = 13.7). Furthermore, the distribution of strong GC connectivity when an airplane is detected is more complex and variable. In the 0-100ms stage, strong connectivity is widely distributed throughout the brain, including the left parastigmal gyrus to the right precuneus of the parietal lobe, the left superior parietal lobe to the isthmus of the right cingulate gyrus (located in the posterior part of the cingulate gyrus), the right superior anterior cingulate gyrus (located in the anterior part of the cingulate gyrus) to the right paracentral lobule, and the right temporal pole to the right olfactory cortex. In the 100-200ms stage, the number of strong connectivity decreases to two, including the left parastigmal gyrus to the right paracentral lobule and the right superior parietal lobe to the right precuneus of the parietal lobe. The duration is 200-300ms, and the number of strong connections is two, including the left insula back to the left superior parietal gyrus and the left temporal olfactory cortex to the right temporal olfactory cortex.

[0078] This embodiment extracts EEG data segments (-20, 300 ms) before and after image stimulation in two scenarios during an infrared background airplane recognition experiment: correct recognition of an airplane (yes1) and correct recognition of no airplane (no1). The aim is to investigate the cognitive process and brain region collaboration mechanisms involved in airplane recognition.

[0079] First, based on time-domain EEG signals from electrodes in four brain regions—anterior, middle, posterior, and whole brain—an SVM classifier was used to decode whether an aircraft was detected. The highest decoding accuracy reached 97.5%, and an accuracy of 81.5% was achieved within 50ms. Comparing the decoding accuracy curves over time based on each region shows that:

[0080] (1) The decoding accuracy curves based on the anterior and posterior brain regions are quite similar to those based on the whole brain. The anterior brain regions mainly include the frontal lobe (F) and a small portion of the central lobe (C), suggesting that the frontal lobe and central lobe may play an important role in the cognitive process of aircraft recognition. The posterior brain regions mainly include the occipital lobe (O), temporal lobe (T), and a small portion of the parietal lobe (P), indicating that the parietal lobe and occipitotemporal lobe also play important roles in the aircraft recognition process. This is consistent with the conclusions of existing research, namely that when the brain recognizes objects (especially man-made objects), not only the visual cortex and higher visual cortex in the occipitotemporal lobe are involved, but also brain regions such as the parietal lobe and frontal lobe.

[0081] (2) The peak of the decoding accuracy curve based on the posterior brain region often appears earlier than other curves, especially in stage one and stage two, indicating that the posterior brain region may be the first to participate in information processing. This may be related to the fact that the occipital-temporal visual cortex located in the posterior brain region is the first to receive visual information.

[0082] (3) The trend of decoding accuracy based on the midbrain region in Stage 1 is similar to that of the whole-brain-based decoding accuracy curve. However, after reaching the highest accuracy (80%) at 110ms, it shows an overall downward trend, which is significantly different from the results based on the whole brain. The midbrain region mainly includes the central area and the parietal lobe. Therefore, it is speculated that the central area and the parietal lobe also participate in the aircraft recognition process, but their contribution is mainly reflected in the short time (0-100ms) when the object is presented. Therefore, in the process of aircraft recognition against an infrared background, multiple parts of the brain, including the anterior, middle, and posterior regions, are involved. Effective aircraft recognition is the result of the joint cooperation of multiple brain regions throughout the whole brain.

[0083] Then, based on the decoding accuracy curve, three stages were defined: 0-100ms, 100-200ms, and 200-300ms. A 68*68 Granger causal network connection matrix was constructed based on the source signals from 68 brain regions of the cerebral cortex obtained through source tracing analysis. Comparing the connection matrices between yes1 and no1 at the same stage reveals that in the connection matrix where no airplane was identified, the stronger GC connections were mainly connections from multiple brain regions to the central region. In the connection matrix where an airplane was correctly identified, not only were the corresponding GC connections from multiple brain regions to the central region stronger, but there were also strong GC connections between other brain regions. Therefore, effective airplane identification may require stronger information exchange and more brain regions to cooperate. Comparing the connection matrices at the three different stages under the same identification conditions shows that as the identification process progresses, GC connections (especially those pointing to the central region) gradually become stronger, with the most significant overall enhancement of the GC network in the 200-300ms stage. An exception is the case of "yes1". Compared to the 100-200ms range, the connections between multiple brain regions pointing to the frontal and parietal lobes are stronger in the connection matrix from 0-100ms. This indicates that the collaboration between different brain regions is constantly evolving during the cognitive process of object recognition. As the cognitive process deepens, information exchange between brain regions tends to become more extensive and active. Regarding the special case of "yes1", combined with the decoding results of "yes1" and "no1" based on whole-brain signals in the 0-100ms stage (accuracy reaching 80% at 50ms), it is speculated that the brain may have already mobilized multiple brain regions to initially and effectively identify the aircraft within the short timeframe of 0-100ms, which is a crucial stage for rapid aircraft recognition. After the information further flows to the central region in the 100-200ms stage, and with more brain regions participating in information processing in the 200-300ms stage, the decoding accuracy of "yes1" and "no1" reaches a maximum of 97.5%, eventually stabilizing at around 90%, thus achieving accurate aircraft recognition.

[0084] Further analysis of the strong GC connections (GC values ​​greater than 95% of the maximum GC value in the connectivity matrix) revealed that the precuneus (involved in memory and environmental perception integration), the parietal lobe (involved in attention and visual perception), and the posterior cingulate cortex and olfactory cortex (involved in memory) all significantly contributed to cognitive activity in both cases. This indicates that aircraft recognition requires not only the visual system but also other brain regions related to memory. The difference lies in the fact that when the aircraft was not recognized, these brain regions acted as starting points in the strong GC connections, directly transmitting information to the paracentral lobule. However, when the aircraft was recognized, there was a richer exchange of information between these brain regions.

[0085] It is noteworthy that, unlike the strong GC connections concentrated in the posterior part of the brain when an aircraft is not recognized, during the 0-100ms period when an aircraft is effectively recognized, the anterior cingulate cortex, located in the front of the brain, also serves as the starting point of a strong GC connection, transmitting information to the paracentral lobule and participating in aircraft recognition. Previous research has shown that the anterior cingulate cortex plays a crucial role in cognitive processes, enabling it to collaborate with surrounding brain regions to complete information processing. Therefore, remote brain regions are indeed involved in object recognition, and may play a vital role in information integration during the short period (0-100ms) of object presentation, supporting the final decision-making process.

[0086] Furthermore, in the cognitive process of the airplane recognition experiment, the paracalar gyrus and paracentral lobule made significant contributions to the recognition activity in most cases. The involvement of the paracalar gyrus may be because the primary visual cortex is mainly located here, making it possible for the left paracalar gyrus to participate in the activity as the starting point of strong GC connections at various stages. The involvement of the paracentral lobule may be due to its association with somatic movement. In this experiment, participants were required to press a button when they recognized an airplane and not press a button when they did not recognize an airplane. Therefore, the paracentral lobule may need to receive visual information from other brain regions to process the recognition results and then make a decision on whether to press a button.

[0087] This embodiment extracts EEG data segments (-20, 300 ms) before and after image stimulation in two scenarios during an infrared background aircraft recognition experiment: correct recognition of an aircraft and correct recognition of no aircraft. The aim is to investigate the cognitive process and brain region collaboration mechanisms involved in aircraft recognition.

[0088] First, based on time-domain EEG signals from electrodes in four brain regions—anterior, middle, posterior, and whole brain—an SVM classifier was used to decode whether an aircraft was detected. The highest decoding accuracy reached 97.5%, with 81.5% accuracy achieved within 50ms. Then, based on the decoding accuracy curve, three stages were defined: 0-100ms, 100-200ms, and 200-300ms. A Granger causal network was constructed based on source signals from 68 brain regions of the cerebral cortex. Comparison revealed that effective aircraft recognition may require more brain regions to participate in stronger information exchange and collaboration, including the paracalcaneal gyrus related to vision and the precuneus, parietal lobe, cingulate gyrus, and olfactory cortex related to memory. Simultaneously, within the short 0-100ms timeframe, the brain mobilized multiple brain regions, including the occipital, temporal, parietal, and cingulate gyrus, to initially and effectively recognize the aircraft; this stage may be crucial for achieving rapid aircraft recognition.

[0089] In summary, this embodiment demonstrates through machine learning decoding and brain network analysis that aircraft recognition requires not only the involvement of the occipital-temporal visual system but also more distant brain regions such as the parietal lobe and cingulate gyrus. Furthermore, the results also prove that the human brain can achieve effective aircraft recognition in a very short time, and its cognitive neural mechanisms are of great significance for subsequently building brain-like models to enable machines to rapidly recognize aircraft.

[0090] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A method for researching brain cognitive mechanism of aircraft identification in infrared background based on electroencephalogram, characterized in that, The method comprises the following steps: Step 1, selecting experimental objects and determining an experimental process; Step 2, pre-processing data obtained through experiments; The step 2 specifically comprises: Firstly, excluding experimental objects with data problems; secondly, removing artifacts caused by fMRI through EGI device supporting software NetStation; and then, using EEGLAB for further pre-processing; The further pre-processing using EEGLAB specifically comprises the following steps: (1) positioning electrodes and deleting four useless electrodes, including E32 and E43 for collecting electrocardiogram and E62 and E63 for collecting electrooculogram; (2) replacing bad leads with poor contact through the average value of signals of surrounding three electrodes; (3) filtering through a 0.3-45Hz FIR band-pass filter, wherein a notch wave is 50Hz; (4) reducing sampling rate to 500Hz to improve calculation speed; (5) re-referencing, converting the reference mode of each channel electroencephalogram to average reference; (6) segmenting, intercepting data segments of correct responses, wherein correct recognition of an airplane and pressing a key are defined as yes1, and correct recognition of no airplane and no pressing of a key are defined as no1, and the picture stimulus before and after the yes1 and no1 is(-20, 300)ms as a target data segment and is subjected to baseline correction; (7) deleting bad segments and identifying and removing artifacts based on ICA, and finally obtaining the number of effective data segments; Step 3, decoding and classifying data obtained through pre-processing through machine learning; The step 3 further comprises: Dividing the whole brain into three parts of front, middle and back, respectively decoding based on multi-channel electroencephalogram time domain signals in the whole brain and three part division zones, to preliminarily explore whether other remote brain areas make contributions to effective recognition of an airplane in addition to the occipital temporal lobe located in the back of the brain during the airplane recognition process of the brain; Step 4, source analysis of electroencephalogram; In the step 4, distributed source model is used to perform source inversion on each segment of data of each subject through a Brainstorm toolbox, and the source inversion specifically comprises the following steps: Firstly, the source space is constrained in the cerebral cortex, and an OpenMEEG BEM algorithm is used to calculate a head model; then, a minimum norm estimation method is used for source estimation, and a sLORETA algorithm is used for normalization; meanwhile, the cerebral cortex is divided into 68 brain regions using a Desikan-Killiany atlas; Step 5, constructing a Granger causality network, and researching the cooperation between brain regions during the airplane recognition cognitive process through construction of a brain network; In the step 5, the 68 brain regions after source inversion are used as network nodes to construct a 68*68 Granger causality network, namely a GC network connection matrix, through a Brainstorm toolbox based on MATLAB.

2. The method of claim 1, wherein the method is used for studying the cognitive mechanism of the brain based on electroencephalogram (EEG) and infrared background down aircraft identification. In the step 1, the experimental process specifically comprises the following steps: An infrared background picture containing an airplane target with different sizes and postures and environmental interference factors is used as a target picture for testing, and an infrared background picture containing only environmental interference factors and no airplane is used as a control picture; when the subject recognizes the target picture containing the airplane, the subject presses the key with the dominant hand, and when the subject recognizes the non-target picture, the subject does not press the key; wherein the ratio of the target picture and the control picture is 1:2, and the pictures are played multiple times with a certain time interval between each group.

3. The method of claim 2, wherein the method is used for studying the cognitive mechanism of the brain based on electroencephalogram (EEG) and infrared background down aircraft identification. During the experiment, the electroencephalogram signal collection adopts the 64-lead electroencephalogram collection equipment of EGI, the electrode distribution is GSN-Hydrocel64, the sampling rate is 1 kHz, and the electrode impedance is maintained below 30KΩ during the collection process.

4. The method of claim 1, wherein the method is used for studying the cognitive mechanism of the brain based on electroencephalogram (EEG) and infrared background down aircraft identification. The step 3 is specifically as follows: The time domain electroencephalogram signals in the yes 1 and no 1 cases are taken as two types of samples, and the model is trained at each time point to classify whether the airplane is recognized; wherein the machine learning model adopts the SVM model provided by Brainstorm, and the model is verified by three-fold cross-validation; in addition, in order to balance the sample amount of the two types, the data segment quantity in the two cases is made the same by random sampling before model training; finally, the model is evaluated according to the classification accuracy.

Images