Neuromorphic visual object classification method based on spiking neural network

Abstract

The present invention relates to the field of visual target classification technology, specifically to a neuromorphic visual target classification method based on a spiking neural network, comprising the following steps: collecting spiking neuron excitation data within a time window, detecting the pulse triggering frequency of target features, analyzing feature correlation fluctuations in visual target classification, identifying feature change trends, screening continuously changing features as stable features, and generating feature change classification results. In the present invention, by adjusting the cross-layer connection paths and synaptic transmission capabilities of neurons, optimization is performed for stable and unstable features, further improving the efficiency and accuracy of visual target classification, especially in dynamic environments. Based on statistical analysis and adjustment of pulse time, the recognition ability of spiking neural networks for time-sensitive targets is enhanced, so that time-aligned pulse data more accurately reflects real-time changes in visual input, significantly optimizing target recognition and classification effects in dynamic scenes.

Description

Technical Field

[0001] The present invention relates to the technical field of visual target classification, and in particular to a neuromorphic visual target classification method based on a spiking neural network. Background Art

[0002] The field of visual object classification encompasses a crucial component of computer vision, focusing on enabling computers to understand and classify different objects in images through image recognition techniques. This field involves a variety of techniques, including image processing, feature extraction, and machine learning algorithms, to automatically identify and classify objects, scenes, or activities. Specific techniques include image preprocessing, feature detection, pattern recognition, and classifier design and optimization, aiming to improve classification accuracy and efficiency through technical means.

[0003] The neuromorphic visual object classification method using spiking neural networks (SNs) uses SNs to classify visual objects. This method mimics the processing methods of biological nervous systems, processing and transmitting information through temporal spiking activity. Technical aspects include the use of specific neuromorphic hardware or simulation software to generate and transmit spiking signals, as well as the target recognition process based on these signals. This method is particularly suitable for processing visual data that is difficult to classify using traditional methods due to its temporal dynamics. Target classification is achieved by designing specific network structures and learning algorithms.

[0004] Existing visual object classification technologies primarily rely on traditional image processing and machine learning algorithms, but they struggle to process dynamic or time-sensitive visual data. Their insensitivity to temporal changes makes it difficult to effectively capture and utilize temporal dynamic information in visual data, especially in rapidly changing scenes. Traditional methods neglect feature stability and instability analysis during feature extraction and classification, resulting in poor recognition accuracy and processing efficiency in complex environments. This limitation hinders widespread deployment in practical applications, particularly in scenarios requiring highly dynamic responses, such as autonomous driving and real-time monitoring. Summary of the Invention

[0005] In order to solve the problems that the existing technology is unable to process dynamic or time-sensitive visual data, lacks sensitivity to time changes, and is difficult to effectively capture and utilize the temporal dynamic information in visual data, especially in rapidly changing scenes. In the process of feature extraction and classification, traditional methods ignore the stability and instability analysis of features, resulting in poor recognition accuracy and processing efficiency in complex environments. The embodiment of the present invention provides a neuromorphic visual target classification method based on a spiking neural network. The technical solution is as follows:

[0006] In one aspect, a neuromorphic visual object classification method based on a spiking neural network is provided, the method comprising:

[0007] S1: Collects pulse neuron excitation data within the time window, detects the pulse trigger frequency of the target feature, analyzes the feature correlation fluctuation of visual target classification, identifies feature change trends, selects continuously changing features as stable features, and generates feature change classification results;

[0008] S2: calling the feature change classification result, analyzing the cross-layer connection paths of neurons with stable features, adjusting synaptic transmission capabilities, avoiding signal transmission delays, screening synaptic connection paths with unstable features, reducing cross-layer connection priorities, and generating cross-layer connection adjustment results;

[0009] S3: Calling the cross-layer connection adjustment result, counting the time distribution of the pulse event, referring to the pulse time pattern of the input sample, analyzing the degree of pulse time deviation, adjusting the time reference point for the dynamic target recognition scenario, evaluating the impact of the time deviation on the pulse time distribution, and obtaining the time reference point adjustment result;

[0010] S4: Using the time reference point adjustment result, according to the pulse trigger time of the input sample, calculate the time scaling ratio, adjust the input data pulse time according to the time scaling ratio, so that the pulse events match the same time scale, and generate time-aligned pulse data.

[0011] As a further solution of the present invention, the feature change classification results include stable feature records, unstable feature records, and feature fluctuation analysis results; the cross-layer connection adjustment results include synaptic transmission adjustment information, connection path priority adjustment results, and signal delay increments; the time reference point adjustment results include time distribution analysis results, time offset evaluation information, and dynamic adjustment conditions; the time-aligned pulse data includes interval consistency, scaling implementation conditions, and time scale matching information.

[0012] As a further solution of the present invention, the steps of obtaining the feature change classification results are specifically as follows:

[0013] S101: Collecting pulse neuron excitation data within a time window, detecting the pulse trigger frequency of the target feature, extracting pulse sequence data, identifying the distribution and change of the trigger frequency of the target feature, and obtaining the target feature pulse response record;

[0014] S102: Based on the target feature impulse response record, analyzing the fluctuation amplitude and change trend of the target feature, classifying the feature change type, screening target features with stable change trends and fluctuation amplitudes within a stable range as stable features, recording target features with unstable change trends and those outside the stable range, and obtaining a target feature change trend identification result;

[0015] S103: calling the target feature change trend identification result, screening the continuously changing target features, recording the fluctuation trend and frequency change, analyzing the fluctuation amplitude of the pulse trigger frequency, and generating a feature change classification result.

[0016] As a further solution of the present invention, the step of obtaining the cross-layer connection adjustment result is specifically:

[0017] S201: Based on the feature change classification results, analyze the connection paths of stable features between differentiated neuron layers, determine whether the synaptic signal matches the signal transmission requirements, optimize the synaptic transmission capacity and increase the path signal transmission time, and obtain the synaptic signal matching result;

[0018] S202: Call the synaptic signal matching result, screen the synaptic connection paths with unstable characteristics, analyze the connection priority of cross-layer transmission, adjust the signal transmission time, weaken the synaptic connection within the signal adjustment range, screen the optimized cross-layer connection paths that meet the requirements, and generate a cross-layer connection adjustment result.

[0019] As a further solution of the present invention, the step of obtaining the time reference point adjustment result is specifically as follows:

[0020] S301: Calling the cross-layer connection adjustment result, counting the time distribution of pulse events, obtaining a time point sequence of multiple pulse events, referring to the pulse time pattern of the input sample, analyzing the distribution of the pulse events in the time dimension, and obtaining the pulse time distribution density;

[0021] S302: Based on the pulse time distribution density, analyze the time span range and the degree of pulse time offset, evaluate the degree of pulse time offset and the degree of time alignment of the input sample time pattern, adjust the time reference point for dynamic target recognition scenarios, match the target motion trajectory, and compare the change in the time alignment error before and after the adjustment to obtain a time reference point error record;

[0022] S303: Calling the time reference point error record, evaluating the impact of the time offset on the pulse time distribution according to the time offset degree of the target motion trajectory, adjusting the time reference point, and generating a time reference point adjustment result.

[0023] As a further solution of the present invention, the evaluation of the degree of pulse time offset and the degree of time alignment of the input sample time pattern is performed using the formula:

[0024] ;

[0025] Among them, D align represents the time alignment error, T pulse,i Represents the time value of the i-th pulse, Tref,i represents the i-th reference time value, W i represents the pulse weight at the i-th time point, and N represents the total number of pulse samples.

[0026] As a further solution of the present invention, the step of acquiring the time-aligned pulse data is specifically as follows:

[0027] S401: calling the time reference point adjustment result, calculating the time interval of the pulse event, extracting the pulse trigger time of the input sample, calculating the mean of the pulse trigger time interval, and obtaining the average time interval of the pulse event;

[0028] S402: Based on the average time interval of the pulse events, calculate the time scaling ratio, adjust the pulse trigger time of the input sample according to the time scaling ratio so that the pulse events match the same time scale, evaluate the distribution consistency of the adjusted pulse time, verify the uniformity of the pulse sequence on the time axis, and obtain time-aligned pulse data.

[0029] As a further solution of the present invention, the formula for adjusting the pulse trigger time of the input sample according to the time scaling ratio is:

[0030] ;

[0031] Among them, t' i represents the adjusted trigger time of the ith pulse, t i represents the original trigger time of the i-th pulse, t min Represents the minimum value of the pulse trigger time, t max represents the maximum value of the pulse trigger time, and τ represents the total duration of the target time scale.

[0032] As a further solution of the present invention, the method further includes step S5:

[0033] S5: Based on the time-aligned pulse data, evaluate the temporal stability of multiple categories of pulse events, calculate the distribution deviation of multiple categories of pulse events within the time window, determine the matching priority of the pulse events, calculate the temporal stability score of the visual target category, adjust the matching degree according to the score, generate a temporal adaptation classification result, optimize the adaptability of the spiking neural network to dynamic visual scenes, and perform target recognition and classification;

[0034] The time adaptation classification result includes the time stability score of the pulse event, the matching priority adjustment result, and the distribution deviation statistical information.

[0035] As a further solution of the present invention, the steps of obtaining the time-adaptive classification results are specifically as follows:

[0036] S501: Based on the time-aligned pulse data, statistics are collected on the distribution of multiple categories of pulse events within a time window, time distribution characteristics of differentiated category pulse events are analyzed, and time deviation values ​​of pulse events are obtained;

[0037] S502: Calling the time deviation values ​​of the multi-category pulse events, analyzing the matching priorities of the multi-category pulse events, calculating the time stability score of the visual target category according to the matching priorities, analyzing the impact of the time stability score on the pulse event matching, adjusting the matching degree, and obtaining a matching time correction result;

[0038] S503: Based on the matching time correction result, the matching method of the pulse event is adjusted, and according to the adjusted pulse event time distribution, a time adaptation classification result is generated to optimize the adaptability of the pulse neural network to dynamic visual scenes, and perform target recognition and classification.

[0039] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least:

[0040] By collecting and analyzing spike neuron excitation data, data that exhibits continuously changing or unstable features within the time window is screened out, effectively improving the dynamic adaptability of feature recognition. By adjusting the cross-layer connection paths and synaptic transmission capabilities of neurons and optimizing for stable and unstable features, the efficiency and accuracy of visual target classification are further improved, especially in dynamic environments. Based on statistical analysis and adjustment of spike times, the spike neural network's ability to recognize time-sensitive targets is enhanced, allowing time-aligned spike data to more accurately reflect real-time changes in visual input. This refined time and feature processing significantly optimizes target recognition and classification in dynamic scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a schematic diagram of the workflow of the present invention; DETAILED DESCRIPTION

[0042] The technical solution of the present invention is described below in conjunction with the accompanying drawings.

[0043] In the embodiments of the present invention, words such as "exemplarily" and "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as an "exemplary" in the present invention should not be interpreted as being preferred or advantageous over other embodiments or designs. Rather, the use of the word "exemplary" is intended to present concepts in a concrete manner. Furthermore, in the embodiments of the present invention, "and / or" can mean both or either of the two.

[0044] In order to make the technical problems, technical solutions and advantages to be solved by the present invention clearer, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0045] See also Figure 1 The embodiment of the present invention provides a neuromorphic visual target classification method based on a spiking neural network. The processing flow of the method may include the following steps:

[0046] S1: Collects spike neuron excitation data within the time window, detects the pulse trigger frequency of the target feature, analyzes the feature correlation fluctuation of visual target classification, identifies feature change trends, selects continuously changing features as stable features, records features with uneven change amplitudes as unstable features, and generates feature change classification results;

[0047] S2: Call the classification results of feature changes, analyze the cross-layer connection paths of neurons with stable features, adjust the synaptic transmission capacity, optimize the pulse propagation efficiency of visual target classification, avoid signal transmission delay, screen the synaptic connection paths with unstable features, reduce the cross-layer connection priority, increase the path signal delay, weaken the synaptic plasticity, and generate the cross-layer connection adjustment results;

[0048] S3: Call the cross-layer connection adjustment results, count the time distribution of the pulse events, refer to the pulse time pattern of the input samples, analyze the time span range and the degree of pulse time deviation, adjust the time reference point for dynamic target recognition scenarios, refer to the time deviation degree of the target motion trajectory, evaluate the impact of time deviation on the pulse time distribution, and obtain the time reference point adjustment result;

[0049] S4: Using the time reference point adjustment result, identify the average interval of the pulse events, calculate the time scaling ratio based on the pulse trigger time of the input sample, adjust the input data pulse time according to the time scaling ratio, so that the pulse events match the same time scale, and generate time-aligned pulse data;

[0050] S5: Based on time-aligned pulse data, evaluate the temporal stability of multi-category pulse events, calculate the distribution deviation of multi-category pulse events within the time window, determine the matching priority of pulse events, calculate the temporal stability score of visual target categories, adjust the matching degree based on the score, generate temporally adapted classification results, optimize the adaptability of the spiking neural network to dynamic visual scenes, and perform target recognition and classification;

[0051] The feature change classification results include stable feature records, unstable feature records, and feature fluctuation analysis results. The cross-layer connection adjustment results include synaptic transmission adjustment information, connection path priority adjustment results, and signal delay increments. The time reference point adjustment results include time distribution analysis results, time offset evaluation information, and dynamic adjustment status. The time-aligned pulse data includes interval consistency, scaling implementation status, and time scale matching information. The time adaptation classification results include the time stability score of the pulse event, matching priority adjustment results, and distribution deviation statistical information.

[0052] The specific steps for obtaining the feature change classification results are as follows:

[0053] S101: Collecting pulse neuron excitation data within a time window, detecting the pulse trigger frequency of the target feature, extracting pulse sequence data, identifying the distribution and change of the trigger frequency of the target feature, and obtaining the target feature pulse response record;

[0054] It is necessary to continuously sample the neuron excitation signal and set the sampling frequency to 10kHz to ensure that all high-frequency pulse signals can be captured. After signal acquisition, the bandpass filtering method is used to remove low-frequency noise and high-frequency interference. After filtering, the data is pulse detected. The pulse triggering frequency is calculated using the counting method, that is, the number of pulses triggered by the neuron is counted within a unit time window (100ms). The calculation formula is:

[0055] ;

[0056] Among them, F t Indicates the pulse trigger frequency within the unit time window, N t is the number of pulses detected in the time window T;

[0057] If 50 pulses are detected within 100ms, then:

[0058] ;

[0059] It is necessary to extract pulse sequence data, store timestamp and pulse amplitude information, use time series analysis method to statistically analyze the trigger frequency distribution of target features, and obtain its mean, variance, kurtosis and other parameters. The mean is used to measure the central trend of pulse triggering, and the variance is used to measure the frequency fluctuation. The threshold is set based on: in the study of biological neuron excitation, a variance less than 50 is set as a stable signal, and a variance greater than 200 is set as a large fluctuation signal. The threshold is based on experimental data statistics and is obtained by calculating the standard deviation of 50 records. If the measured variance is greater than this value, it is necessary to further analyze its changing trend. The pulse trigger frequency of a certain neuron target feature fluctuates around 500, and the measured standard deviation is 8. If it is lower than 50, it can be judged as a stable signal. If the standard deviation of a feature reaches 220, it is judged that the signal fluctuates greatly, forming a target feature pulse response record.

[0060] S102: Based on the target feature pulse response record, analyze the fluctuation amplitude and change trend of the target feature, classify the feature change type, select target features with stable change trends and fluctuation amplitudes within a stable range and classify them as stable features, record target features with unstable change trends and those outside the stable range, and obtain target feature change trend identification results;

[0061] Analyze the fluctuation amplitude and change trend of the target characteristics and calculate the fluctuation amplitude, that is, the difference between the maximum and minimum values ​​of the pulse trigger frequency. The calculation formula is:

[0062] ;

[0063] Among them, the parameter F max is the maximum pulse frequency in the time window, F min is the minimum pulse frequency;

[0064] If the pulse frequency of a target feature has a maximum value of 800 Hz and a minimum value of 500 Hz within 100 ms, then its fluctuation amplitude is:

[0065] ;

[0066] Identify the changing trend and use the sliding average method to perform trend analysis, calculating the average of the first five windows, and so on;

[0067] Stable interval threshold setting: In bioelectric signal processing, the stable interval is set within the range of ±10% of the mean value of the target feature, that is, if the mean value of the target feature is 600Hz, the stable interval is set to 540Hz to 660Hz. If the fluctuation exceeds this range, it is judged as an unstable feature. If the frequency mean of a target feature is 550Hz and the measured data are [510, 520, 530, 515, 525]Hz, all the data are in the stable interval (495Hz to 605Hz), so the feature can be judged as a stable feature. If the measured data is [510, 650, 490, 700, 450]Hz, it exceeds the stable interval range and further processing is required to obtain the target feature change trend identification result.

[0068] S103: Calling the target feature change trend identification result, screening the continuously changing target features, recording the fluctuation trend and frequency change, analyzing the fluctuation amplitude of the pulse trigger frequency, and generating the feature change classification result;

[0069] Filter the continuously changing target features, traverse the pulse trigger frequency sequences of all target features, extract the features that continuously change in multiple time windows, record the frequency data of the features, identify the fluctuation trend of each feature, determine the overall direction of frequency increase or decrease, analyze the amplitude of frequency change, observe the range of change by calculating the difference between each time point, and evaluate the rate of change. If the fluctuation range of a target feature is large and fails to stabilize for a long time, further analyze its frequency change pattern. If the trigger frequency of a feature shows an increasing trend in multiple consecutive time windows and its increase rate continues to expand, it can be judged that the feature has a significant growth trend. If a feature shows frequent ups and downs in different time windows, it is necessary to further judge whether there is periodic change. By comparing the changes of multiple target features, the feature change classification results are generated.

[0070] The specific steps for obtaining the cross-layer connection adjustment results are:

[0071] S201: Based on the feature change classification results, analyze the connection paths of stable features between differentiated neuron layers, determine whether the synaptic signal matches the signal transmission requirements, optimize the synaptic transmission capacity and increase the path signal transmission time, and obtain the synaptic signal matching result;

[0072] Analyze the connection paths of stable features between differentiated neuron layers, construct the topological structure of the neuron network, determine the neurons to which the stable features belong and their synaptic connections, and use the adjacency matrix A to represent the connection path of each neuron node. Let A(i, j) = 1 indicate that there is a synaptic connection between neurons i and j, and A(i, j) = 0 indicates no connection. Further calculate the synaptic signal transmission delay. Assuming that the synaptic delay time from neuron i to j is T(i, j), the calculation formula is:

[0073] ;

[0074] Where d(i,j) is the synaptic distance between neurons i and j, V s is the signal transmission speed;

[0075] If d(i,j)=0.2mm, V s =1.5m / s, then:

[0076] ;

[0077] Determine whether the synaptic signal matches the signal transmission requirement and calculate the matching error E_m of the synaptic signal.

[0078] Synaptic signal matching error setting: The error range of the synaptic delay time is set within ±10% of the average transmission time of the target neuron. That is, if the average transmission time of the neuron is 1.2ms, the synaptic signal matching error is set between 1.08ms and 1.32ms. If the error exceeds this range, the synaptic signal is considered unmatched. If the average transmission time of a neuron is 1.5ms, but the calculated synaptic delay time is 2.0ms, the error is 0.5ms, which exceeds the matching error range. In this case, the synaptic transmission capacity needs to be optimized. The optimization method includes adjusting the synaptic weighting coefficient to improve signal transmission efficiency, or adding new connection paths to reduce transmission time to obtain the synaptic signal matching result.

[0079] S202: Calling the synaptic signal matching results, screening synaptic connection paths with unstable characteristics, analyzing the connection priority of cross-layer transmission, adjusting the signal transmission time, weakening the synaptic connections within the signal adjustment range, screening the optimized cross-layer connection paths that meet the requirements, and generating a cross-layer connection adjustment result;

[0080] Screen the synaptic connection paths with unstable characteristics, extract all neurons with unstable characteristics, and determine their cross-layer transmission paths based on the topological structure A. Calculate the priority P(i, j) of cross-layer transmission and set the calculation formula as:

[0081] ;

[0082] Among them, W syn(i, j) is the synaptic weight from neuron i to j, T(i, j) is the synaptic signal transmission time;

[0083] The higher the priority, the more important the connection path is in cross-layer transmission;

[0084] If W syn (i,j)=0.8, T(i,j)=1.2ms, then P(i,j)=0.67;

[0085] If W syn (i,j)=0.6, T(i,k)=1.8ms, then P(i,k)=0.33;

[0086] Therefore, the priority of connection (i, j) is higher than that of connection (i, k);

[0087] Priority threshold setting: According to the adaptive mechanism of biological neural networks, the threshold of cross-layer transmission priority is set to 0.5. If the priority of the synaptic connection path is lower than this value, the signal transmission time is adjusted to reduce its impact by weakening the synaptic connection weight. If the priority of a synaptic connection path is 0.3, its synaptic weight is weakened to 0.1 to reduce its impact. After screening and optimization, the cross-layer connection path that meets the requirements is generated to generate the cross-layer connection adjustment result.

[0088] The specific steps for obtaining the time reference point adjustment result are:

[0089] S301: Call the cross-layer connection adjustment result, count the time distribution of the pulse event, obtain the time point sequence of multiple pulse events, refer to the pulse time pattern of the input sample, analyze the distribution of the pulse event in the time dimension, and obtain the pulse time distribution density;

[0090] Perform time statistics on pulse events, extract the occurrence time of pulse events, and form a time series. For this time series, set a time window, for example, count the number of pulse events in a time window of 100 milliseconds, and calculate the density of pulse events per unit time. For the distribution of pulse events, use a histogram to count the number of pulses in different time periods, and use the Gaussian distribution fitting method to perform distribution modeling, thereby forming a time distribution curve of pulse events. In a certain time window, if 50 pulse events are counted, and the average occurrence frequency of the overall pulse events is 40 times per 100 milliseconds, then the pulse density in the window is higher than the overall mean. It can be judged that some time periods are high-incidence areas of pulse events. At the same time, refer to the pulse time pattern of the input sample, compare and analyze the sample pulse sequence with the statistically obtained pulse distribution density, set the time step, so that the sample pulse time pattern slides and aligns with the statistical distribution, and calculates the pulse event matching degree at each time step. If an input sample is in the range of 0 to 100 milliseconds The sample contains 40 pulse events, while the number of pulse events in the same time interval in the statistical data is 45. The matching degree is calculated as 40 divided by 45, which is approximately 0.89, thus obtaining the matching degree between the sample pattern and the statistical data. In this process, to ensure that the statistical density of the pulse events meets the standards of the target recognition device, a pulse density threshold needs to be set. This threshold is calculated based on the target device's requirements for the pulse event frequency. If the target recognition device requires a minimum resolvable time interval of 5 milliseconds, the device requires at least 200 pulse events per second. Therefore, the pulse density threshold is set to 200 times per second. If the pulse density in a certain time window is lower than the threshold, it is considered a pulse sparse area and does not participate in the pattern matching calculation. If the minimum sampling interval of the device is set to 10 microseconds, the density threshold can be further refined to 100,000 times per second to ensure that the calculation result is within the processing range of the device. The threshold is adjusted based on the matching between the input sample and the pulse time distribution density to obtain the pulse time distribution density.

[0091] S302: Based on the pulse time distribution density, the time span range and the degree of pulse time deviation are analyzed, and the degree of time alignment between the pulse time deviation and the input sample time pattern is evaluated. For dynamic target recognition scenarios, the time reference point is adjusted to match the target motion trajectory. The change in the time alignment error before and after the adjustment is compared to obtain a time reference point error record.

[0092] The degree of pulse time deviation and time alignment of the input sample time pattern is evaluated using the formula:

[0093] ;

[0094] Among them, D align represents the time alignment error, T pulse,i Represents the time value of the i-th pulse, Tref,i represents the i-th reference time value, W i represents the pulse weight at the i-th time point, and N represents the total number of pulse samples;

[0095] Parameter interpretation and calculation process:

[0096] Pulse time value T pulse,i :The time point of the i-th pulse arrival is recorded by a high-precision time measurement device. In one measurement, the recorded pulse time value is:

[0097] ;

[0098] Reference time value T ref,i : According to the designed or expected pulse arrival time, set the reference time value of the i-th pulse. Corresponding to the above pulse time, the reference time value is:

[0099] ;

[0100] Pulse weight W i : Set the weight according to the importance or signal strength of each pulse. The weight setting basis can be factors such as signal strength and signal-to-noise ratio. The weight value obtained through measurement is:

[0101] ;

[0102] Calculate the absolute value of the weighted time deviation for each pulse:

[0103] For the first pulse:

[0104] ;

[0105] For the second pulse:

[0106] ;

[0107] ;

[0108] For the third pulse:

[0109] ;

[0110] Compute the sum of the absolute values ​​of the weighted time deviations:

[0111] ;

[0112] Compute the square root of the sum of the squared weights:

[0113] ;

[0114] Calculate the time alignment error value:

[0115] ;

[0116] The result shows that after calculation, the time alignment error value is about 0.000639 seconds. This value reflects the weighted average deviation between the pulse time and the reference time. The smaller the value, the higher the degree of alignment between the pulse time and the reference time.

[0117] S303: Calling the time reference point error record, evaluating the impact of the time offset on the pulse time distribution based on the time offset degree of the target motion trajectory, adjusting the time reference point, and generating a time reference point adjustment result;

[0118] Based on the degree of time offset of the target's motion trajectory, compare the pulse time distribution before and after adjustment. Calculate the mean of the pulse time distribution before and after adjustment. If the mean before adjustment is 58 milliseconds and the mean after adjustment is 62 milliseconds, calculate the offset of the pulse time distribution, which is equal to the mean after adjustment minus the mean before adjustment, which is 4 milliseconds. To assess the impact of time offset on the pulse time distribution, calculate the mean square error before and after adjustment. For example, if the mean square error before adjustment is 10 milliseconds squared and the mean square error after adjustment is 6 milliseconds squared, the calculated error reduction is 4 milliseconds squared. To ensure the rationality of the adjustment, set a mean square error threshold based on the device's acceptable range for time alignment error. If the maximum acceptable alignment error for the device is 5 milliseconds squared, set the error threshold to 5 milliseconds squared. If the calculated error reduction is still above the threshold, further optimize the time reference point. If the mean square error after adjustment is below the threshold, or if the calculated result is 3 milliseconds squared, the time alignment is considered acceptable and no further time reference point adjustment is required. Generate the time reference point adjustment result.

[0119] The specific steps for acquiring time-aligned pulse data are as follows:

[0120] S401: Call the time reference point adjustment result, calculate the time interval of the pulse event, extract the pulse trigger time of the input sample, calculate the average of the pulse trigger time interval, and obtain the average time interval of the pulse event;

[0121] Calculate the time interval of pulse events, extract the pulse trigger time of the input samples, and arrange them in chronological order, calculate the time interval between adjacent pulse events, if the pulse trigger time of an input sample is 10 milliseconds, 25 milliseconds, 45 milliseconds, and 70 milliseconds, then the adjacent time intervals are 15 milliseconds, 20 milliseconds, and 25 milliseconds respectively, calculate the average of the pulse trigger time interval, which is defined as the sum of all time intervals divided by the number of intervals. The average time interval of the above data is calculated as 60 milliseconds divided by 3, which is 20 milliseconds. In this process, in order to ensure that the interval calculation of the pulse event meets the time resolution requirements of the device, it is necessary to set the time interval. The interval baseline value is set based on the device's minimum sampling period or pulse triggering mechanism. In a neuromorphic computing device, the time resolution is set to 1 millisecond, so the time interval baseline value should be set to an integer multiple of 1 millisecond, such as 10 milliseconds, 20 milliseconds, etc. If the calculated average time interval of a pulse event is 19 milliseconds, the device will automatically adjust to the closest baseline value of 20 milliseconds to meet the device timing requirements. The baseline value setting can be optimized according to the device operating environment. In a high-load environment, the time interval baseline value needs to be increased to 50 milliseconds to reduce the calculation overhead of the pulse event and obtain the average time interval of the pulse event.

[0122] S402: Calculate the time scaling ratio based on the average time interval of the pulse events, adjust the pulse trigger time of the input sample according to the time scaling ratio so that the pulse events match the same time scale, evaluate the distribution consistency of the adjusted pulse time, verify the uniformity of the pulse sequence on the time axis, and obtain time-aligned pulse data;

[0123] The formula for adjusting the pulse trigger time of the input sample by time scaling is:

[0124] ;

[0125] Among them, t' i represents the adjusted trigger time of the ith pulse, t i represents the original trigger time of the i-th pulse, t min Represents the minimum value of the pulse trigger time, t max represents the maximum value of the pulse triggering time, and τ represents the total duration of the target time scale;

[0126] Parameter meaning and calculation process:

[0127] Parameter t i represents the original trigger time of the i-th pulse, which is directly obtained from the pulse event data monitoring device;

[0128] t min and t maxare the minimum and maximum values ​​of all pulse triggering times, respectively, which are calculated by counting the time points in the pulse data set;

[0129] τ is the total duration of the target time scale, which is determined according to the experimental design requirements and reflects the length of the time frame to be adjusted;

[0130] Taking actual data as an example, a set of pulse event time data is obtained from a sensor network, and the following time points are monitored (unit: milliseconds): {100, 150, 200, 250, 300};

[0131] For this set of data, calculate t min =100 and t max = 300, set the target time scale τ = 1 second (1000 milliseconds);

[0132] Calculate the adjusted time of the first pulse:

[0133] ;

[0134] Calculate the adjusted time of the second pulse:

[0135] ;

[0136] Calculate the adjusted time of the third pulse:

[0137] ;

[0138] Calculate the adjusted time of the fourth pulse:

[0139] ;

[0140] Calculate the adjusted time of the fifth pulse:

[0141] ;

[0142] The above calculation results show how to adjust the original pulse trigger time to different time points within 1 second. The results show that the adjusted pulse time can be evenly distributed on the set time scale, which meets the requirement of matching pulse events to the same time scale. It helps to further analyze the distribution consistency of pulse time and the uniformity of the sequence. The adjusted time points will be used to evaluate the consistency of the pulse time distribution and verify the uniformity of the pulse data after time alignment.

[0143] The specific steps for obtaining the time-adaptive classification results are as follows:

[0144] S501: Based on the time-aligned pulse data, count the distribution of multiple categories of pulse events in the time window, analyze the time distribution characteristics of differentiated category pulse events, and obtain time deviation values ​​of the multiple categories of pulse events;

[0145] To perform statistics on pulse events of different categories within a time window, it is necessary to set a fixed time window of 100 milliseconds and segment all pulse events according to the window to ensure that all events can be counted within the same time scale. In each time window, the number of pulse events of different categories is counted to form time series data. In the first time window, pulse events of category A occur 15 times, category B occurs 8 times, and category C occurs 20 times. The data of this window is recorded as [15, 8, 20]. After the data of all windows are accumulated, a time series data is formed. The distribution matrix requires calculating the time deviation values ​​of different categories of pulse events. The time deviation values ​​can be obtained by calculating the fluctuation of the category in all time windows. The number of pulse events of category A in 10 time windows is 10, 12, 8, 15, 11, 14, 9, 13, 10, and 12, respectively. The overall fluctuation can be calculated to obtain the time deviation value of the category. The time deviation values ​​of the remaining categories can be calculated in the same way and normalized so that the time deviation values ​​of all categories are within the same magnitude range. After the calculation is completed, the time deviation values ​​of the pulse events are obtained.

[0146] S502: Calling the time deviation values ​​of multiple categories of pulse events, analyzing the matching priorities of the multiple categories of pulse events, calculating the temporal stability score of the visual target category based on the matching priorities, analyzing the impact of the temporal stability score on the matching of the pulse events, adjusting the matching degree, and obtaining the matching time correction result;

[0147] The matching priority of each category of pulse events is analyzed, and the priority weight is determined according to the time deviation value. The weight setting can adopt an inverse relationship, that is, the smaller the time deviation value of the category, the greater its priority weight. If the standardized time deviation value of a category is set to 0.2, its matching priority weight is high, while the time deviation value of another category is 0.5, its matching priority is low. The temporal stability score of the visual target category is calculated. The score is related to the time deviation value and the priority weight. If the priority weight of a category is high but the time deviation value is still large, the temporal stability score of the category will be affected to a certain extent. After the temporal stability score is calculated, an impact analysis is required, and an impact threshold is set to 0.6. If the score of a category is lower than this value, its matching degree is reduced, and the matching method is adjusted to weaken the matching degree by a certain proportion, so that the matching priority of the category is reduced. After the adjustment is completed, the matching time correction result is obtained, and the category matching priority is reallocated according to the score after matching adjustment to obtain the matching time correction result.

[0148] S503: Based on the matching time correction result, the matching method of the pulse event is adjusted, and according to the adjusted pulse event time distribution, a time adaptation classification result is generated to optimize the adaptability of the spiking neural network to dynamic visual scenes, and perform target recognition and classification;

[0149] The matching method of pulse events is adjusted, and a time distribution adaptation matrix is ​​established. This matrix is ​​used to redistribute the time matching ratio of pulse events. If the matching score of a certain category is high, the number of pulse events in this category will increase within a certain range in the final time adaptation classification. If the matching scores of categories A, B, and C are 0.7, 0.5, and 0.9, respectively, their time distribution adaptation results will tilt towards the category with the high matching score. The pulse event matching method is adjusted according to this matrix. If the original event number matrix is ​​[15, 8, 20], it will become [13, 9, 21] after adjustment, so that the event distribution is more in line with the matching priority requirements. Based on the adjusted time distribution of pulse events, the time adaptation classification results are generated, and the time adaptation parameters of the spiking neural network are optimized. The optimization method can adjust the matching degree through normalization so that the matching coefficient of each category is within a reasonable range. If the matching coefficients are 0.8, 0.75, and 0.9, respectively, they will be adjusted to [0.84, 0.79, 0.91] after optimization, completing the matching adjustment and classification optimization of pulse events.

[0150] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A neuromorphic visual object classification method based on a spiking neural network, characterized in that: The following steps are involved: S1: Collects pulse neuron excitation data within the time window, detects the pulse trigger frequency of the target feature, analyzes the feature correlation fluctuation of visual target classification, identifies feature change trends, selects continuously changing features as stable features, and generates feature change classification results; S2: calling the feature change classification result, analyzing the cross-layer connection paths of neurons with stable features, adjusting synaptic transmission capabilities, avoiding signal transmission delays, screening synaptic connection paths with unstable features, reducing cross-layer connection priorities, and generating cross-layer connection adjustment results; S3: Calling the cross-layer connection adjustment result, counting the time distribution of the pulse event, referring to the pulse time pattern of the input sample, analyzing the degree of pulse time deviation, adjusting the time reference point for the dynamic target recognition scenario, evaluating the impact of the time deviation on the pulse time distribution, and obtaining the time reference point adjustment result; S4: using the time reference point adjustment result, calculating a time scaling ratio according to the pulse trigger time of the input sample, adjusting the input data pulse time according to the time scaling ratio so that the pulse events match the same time scale, and generating time-aligned pulse data; S5: Based on the time-aligned pulse data, evaluate the temporal stability of multiple categories of pulse events, calculate the distribution deviation of multiple categories of pulse events within the time window, determine the matching priority of the pulse events, calculate the temporal stability score of the visual target category, adjust the matching degree according to the score, generate a temporal adaptation classification result, optimize the adaptability of the spiking neural network to dynamic visual scenes, and perform target recognition and classification; The time adaptation classification result includes the time stability score of the pulse event, the matching priority adjustment result, and the distribution deviation statistical information.

2. The neuromorphic visual target classification method based on spiking neural network according to claim 1 is characterized in that: The feature change classification results include stable feature records, unstable feature records, and feature fluctuation analysis results; the cross-layer connection adjustment results include synaptic transmission adjustment information, connection path priority adjustment results, and signal delay increments; the time reference point adjustment results include time distribution analysis results, time offset evaluation information, and dynamic adjustment conditions; the time-aligned pulse data includes interval consistency, scaling implementation conditions, and time scale matching information.

3. The neuromorphic visual target classification method based on spiking neural network according to claim 1, characterized in that: The steps for obtaining the feature change classification result are specifically as follows: S101: Collecting pulse neuron excitation data within a time window, detecting the pulse trigger frequency of the target feature, extracting pulse sequence data, identifying the distribution and change of the trigger frequency of the target feature, and obtaining the target feature pulse response record; S102: Based on the target feature impulse response record, analyzing the fluctuation amplitude and change trend of the target feature, classifying the feature change type, screening target features with stable change trends and fluctuation amplitudes within a stable range as stable features, recording target features with unstable change trends and those outside the stable range, and obtaining a target feature change trend identification result; S103: calling the target feature change trend identification result, screening the continuously changing target features, recording the fluctuation trend and frequency change, analyzing the fluctuation amplitude of the pulse trigger frequency, and generating a feature change classification result.

4. The neuromorphic visual target classification method based on spiking neural network according to claim 3 is characterized in that: The steps for obtaining the cross-layer connection adjustment result are specifically as follows: S201: Based on the feature change classification results, analyze the connection paths of stable features between differentiated neuron layers, determine whether the synaptic signal matches the signal transmission requirements, optimize the synaptic transmission capacity and increase the path signal transmission time, and obtain the synaptic signal matching result; S202: Call the synaptic signal matching result, screen the synaptic connection paths with unstable characteristics, analyze the connection priority of cross-layer transmission, adjust the signal transmission time, weaken the synaptic connection within the signal adjustment range, screen the optimized cross-layer connection paths that meet the requirements, and generate a cross-layer connection adjustment result.

5. The neuromorphic visual target classification method based on spiking neural network according to claim 4 is characterized in that: The steps for obtaining the time reference point adjustment result are specifically as follows: S301: Calling the cross-layer connection adjustment result, counting the time distribution of pulse events, obtaining a time point sequence of multiple pulse events, referring to the pulse time pattern of the input sample, analyzing the distribution of the pulse events in the time dimension, and obtaining the pulse time distribution density; S302: Based on the pulse time distribution density, analyze the time span range and the degree of pulse time offset, evaluate the degree of pulse time offset and the degree of time alignment of the input sample time pattern, adjust the time reference point for dynamic target recognition scenarios, match the target motion trajectory, and compare the change in the time alignment error before and after the adjustment to obtain a time reference point error record; S303: Calling the time reference point error record, evaluating the impact of the time offset on the pulse time distribution according to the time offset degree of the target motion trajectory, adjusting the time reference point, and generating a time reference point adjustment result.

6. The neuromorphic visual target classification method based on spiking neural network according to claim 4, characterized in that: The evaluation of the degree of pulse time deviation and the degree of time alignment of the input sample time pattern is performed using the formula: ； in, represents the time alignment error, Representative Pulse time value, Representative A reference time value, Representative The pulse weight at each time point, Represents the total number of pulse samples.

7. The neuromorphic visual target classification method based on spiking neural network according to claim 5, characterized in that: The steps for acquiring the time-aligned pulse data are specifically as follows: S401: calling the time reference point adjustment result, calculating the time interval of the pulse event, extracting the pulse trigger time of the input sample, calculating the mean of the pulse trigger time interval, and obtaining the average time interval of the pulse event; S402: Based on the average time interval of the pulse events, calculate the time scaling ratio, adjust the pulse trigger time of the input sample according to the time scaling ratio so that the pulse events match the same time scale, evaluate the distribution consistency of the adjusted pulse time, verify the uniformity of the pulse sequence on the time axis, and obtain time-aligned pulse data.

8. The neuromorphic visual target classification method based on spiking neural network according to claim 7, characterized in that: The formula for adjusting the pulse trigger time of the input sample according to the time scaling ratio is: ； in, Representative The adjusted trigger time of each pulse, represent The original trigger time of the pulse, Represents the minimum value of the pulse trigger time, Represents the maximum value of the pulse trigger time, Represents the total duration of the target time scale.

9. The neuromorphic visual target classification method based on spiking neural network according to claim 7, characterized in that: The steps for obtaining the time adaptation classification result are specifically as follows: S501: Based on the time-aligned pulse data, statistics are collected on the distribution of multiple categories of pulse events within the time window, time distribution characteristics of differentiated category pulse events are analyzed, and time deviation values ​​of the multiple categories of pulse events are obtained; S502: Calling the time deviation values ​​of the multi-category pulse events, analyzing the matching priorities of the multi-category pulse events, calculating the time stability score of the visual target category according to the matching priorities, analyzing the impact of the time stability score on the pulse event matching, adjusting the matching degree, and obtaining a matching time correction result; S503: Based on the matching time correction result, the matching method of the pulse event is adjusted, and according to the adjusted pulse event time distribution, a time adaptation classification result is generated to optimize the adaptability of the pulse neural network to dynamic visual scenes, and perform target recognition and classification.

Images