WO2021143066A1 - Target recognition method, device, and system, and computer readable storage medium - Google Patents

Abstract

Disclosed in the present invention are a target recognition method, device, and system, and a computer readable storage medium. The method comprises the following steps: obtaining original pulse data; determining a pulse sampling window; inputting pulse in the pulse sampling window into a pulse neural network; mapping the pulse that is inputted into the pulse neural network; in the case that the conditions are satisfied, sequentially transmitting the pulse along each layer of excitation neurons excluding a last layer of excitation neurons, transmitting the pulse to an inference layer along the last layer of excitation neurons, and sequentially transmitting the pulse along each layer of inference neurons excluding the last layer of inference neurons; and determining a recognition result. The device comprises a pulse obtaining module, a sampling window module, a pulse mapping module, a neuron excitation module, a neuron inferencing module, and a recognition result determining module. The system comprises the foregoing device. The present invention can achieve the accurate and quick recognition of a target to be recognized, can be well suitable for a target having a high movement speed, and can consider recognition accuracy and the amount of calculation.

Description

Target recognition method, device, system and computer readable storage medium Technical field

The present invention relates to the technical field of target recognition. More specifically, the present invention is a target recognition method, device, system, and computer-readable storage medium.

Background technique

In recent years, the field of artificial intelligence has developed rapidly, with breakthroughs in algorithms, hardware, chips, etc., especially in the field of imagery represented by computer vision, and it has widely commercialized and civilianized many research results, which is extremely convenient. To improve people’s daily life. However, the current artificial intelligence algorithms still take the second-generation artificial neural network as the main core. This type of algorithm was proposed as early as 80 years in the last century. With years of continuous research, the current academic community has fallen into a research bottleneck. Therefore, people have shifted the focus to the third-generation artificial neural network (ie, spiking neural network). However, due to the limitations of the existing technology, for computer vision applications, especially target recognition, there are often low recognition accuracy and slow recognition speed. , The amount of calculation is large, the hardware requirements are too high, and the power consumption is too large.

Therefore, how to effectively improve the accuracy and speed of target recognition, and reduce the requirements for calculation and power consumption of target recognition, has become a technical problem to be solved urgently and a focus of research by those skilled in the art.

Summary of the invention

In order to solve the problems of poor recognition effect and the inability to effectively recognize high-speed moving targets in the existing target recognition technology, the present invention innovatively provides a target recognition method, device, system and computer-readable storage medium. The neural network performs noise filtering and pulse enhancement on the original input pulse, which can achieve accurate and efficient identification of the target to be identified, and can also significantly reduce the amount of calculation and power consumption.

In order to achieve the above technical purpose, the present invention discloses a target recognition method, and the method includes the following steps:

Acquiring raw pulse data of the photographed target to be identified, where the raw pulse data includes at least one pulse sequence;

The area where the target to be identified is located is retrieved by screening each pulse in the pulse sequence, the pulse sampling window is determined according to the area where the target to be identified is located, and the pulses in the pulse sampling window are input to the pulse neural network, the pulse neural network Including excitation layer and reasoning layer;

Mapping the pulses input to the spiking neural network so that the number of firing neurons in the first layer of the firing layer corresponds to the size of the pulse sampling window, and the firing layer includes multiple layers of firing neurons connected in sequence;

When the first preset condition is met, the pulse is transmitted along the firing neurons of each layer other than the final layer in turn, and the pulse is transmitted to the inference layer along the firing neuron of the last layer when the second preset condition is met, the reasoning layer Contains multiple layers of inference neurons connected in sequence;

When the third preset condition is satisfied, the pulse is transmitted in sequence along each layer of inference neurons except the last layer of inference neurons;

The recognition result is determined according to the activity of the last layer of inference neurons.

Further, the acquired raw pulse data of the target to be recognized comes from a bionic vision sensor, and the bionic vision sensor is used to photograph the target to be recognized.

Further, the first preset condition is that the change in membrane potential of each layer of stimulated neurons other than the final layer of stimulated neurons exceeds its own first excitation threshold.

Further, the second preset condition is that the change in membrane potential of the final layer of stimulated neurons exceeds its own second threshold for excitation, and the proportion of the number of activated neurons in the final layer of stimulated neurons is greater than the first preset Proportion.

Further, the third preset condition is that the change in membrane potential of each layer of inference neurons other than the last layer of inference neurons exceeds its own third excitation threshold.

Further, after the pulse sampling window is determined, it further includes the step of recording the time interval when the target to be identified appears in the pulse sampling window twice.

Further, the original pulse data includes multiple pulse sequences, and the recognition results are output only when the recognition results of all the pulse sequences after the above-mentioned processing are consistent.

In order to achieve the above technical purpose, the present invention also discloses a target recognition device, and the device includes a pulse acquisition module, a sampling window module, a pulse mapping module, a neuron excitation module, a neuron inference module, and a recognition result determination module;

The pulse acquisition module is used to acquire the original pulse data of the target to be identified that is photographed; the original pulse data includes at least one pulse sequence;

The sampling window module is used to retrieve the area where the target to be recognized is located by filtering each pulse in the pulse sequence, and is used to determine the pulse sampling window according to the area where the target to be recognized is located, and to sample the pulse in the window The pulses of are input to the impulse neural network; the impulse neural network includes an excitation layer and an inference layer;

The pulse mapping module is used to map the pulses input to the pulse neural network so that the number of the first layer of excitation neurons in the excitation layer corresponds to the size of the pulse sampling window; the excitation layer contains multiple layers of sequentially connected excitation nerves Yuan;

The neuron excitation module is used to enable the pulse to be transmitted along the firing neurons in each layer other than the terminal layer when the first preset condition is satisfied, and to enable the pulse to fire the nerve along the terminal layer when the second preset condition is satisfied Elements are transferred to the inference layer; the inference layer includes multiple layers of inference neurons connected in sequence;

The neuron inference module is used to enable pulses to be sequentially transmitted along each layer of inference neurons other than the last layer of inference neurons when the third preset condition is satisfied;

The recognition result determination module is used to determine the recognition result according to the activity of the last layer of inference neurons.

In order to achieve the above-mentioned technical purpose, the present invention also discloses a target recognition system, which includes a bionic vision sensor and the above-mentioned target recognition device. The bionic vision sensor is used to photograph the target to be recognized to obtain the target to be recognized. The raw pulse data.

In order to achieve the above technical objectives, the present invention also discloses a computer-readable storage medium on which a computer program is stored, and the computer program is executed by a processor to implement any of the above-mentioned target identification methods.

The beneficial effects of the present invention are as follows: Compared with the prior art, the present invention can realize the accurate and fast recognition of the target to be recognized, and can be better applied to the target with high moving speed. The present invention can take into account the accurate recognition at the same time. The problem with the amount of calculation and power consumption. The invention can also realize the measurement of the rotation speed of the high-speed rotating target, thereby providing a better solution for the rotation speed calibration of the rotating body.

The invention can record the dynamic change information of the high-speed moving target through the bionic vision sensor with high time resolution, and process the pulse sequence output by the bionic vision sensor through the pulse neural network, and can realize the noise filtering and pulse enhancement of the original input pulse to compare To achieve accurate and rapid identification of the target, specifically, the pulse input of the inference layer can be reduced by controlling the activation ratio of the pulse neuron in the excitation layer, thereby reducing the calculation amount and power consumption of the entire pulse processing algorithm; The secondary measurement mechanism can significantly improve the recognition accuracy of the pulsed neural network; the present invention realizes the rotation speed measurement of the high-speed rotating target by analyzing the characteristics of the circular motion and adopts a fixed sampling window method, thereby providing a realization method for the rotation speed calibration of the rotating body.

Description of the drawings

FIG. 1 is a schematic flowchart of a target recognition method in some embodiments of the present invention.

Figure 2 is a schematic diagram of the composition of a target recognition system in some embodiments of the present invention.

FIG. 3 is a schematic diagram of the working principle of the pulse recognition process in some embodiments of the present invention taking a plurality of stimulated neurons with a local connection relationship as an example.

FIG. 4 is a schematic diagram of the working principle of the pulse recognition process in other embodiments of the present invention.

Figure 5 shows the effect of filtering and enhancing the pulse sequence of the region where the character "P" is located, which is the target to be recognized through the excitation layer.

Figure 6 shows the effect of noise filtering and pulse enhancement on the pulse sequence of the region where the target to be recognized—the character "C" is located through the excitation layer.

Fig. 7 shows the effect of noise filtering and pulse enhancement on the pulse sequence of the region where the target to be recognized—the character "L" is located through the excitation layer.

Detailed ways

The following is a detailed explanation and description of a target recognition method, device, system, and computer-readable storage medium provided by the present invention with reference to the accompanying drawings of the specification.

Please refer to FIG. 1. FIG. 1 is a schematic flowchart of a target recognition method in some embodiments of the present invention. This embodiment provides a target recognition method. Specifically, the target recognition method may include the following steps.

Step 100: In this embodiment, the target to be recognized can be photographed by the bionic vision sensor to obtain the original pulse data of the target to be recognized. The bionic vision sensor used is an electronic device manufactured by imitating the vision sensing principle of the eyes of a living body. To provide information input for subsequent brain-like neural network algorithms, bionic vision sensors often have high time resolution and can record dynamic change information of high-speed moving targets. The bionic vision sensors used in the present invention include but are not limited to the following three types : Differential sensor, integral sensor and event sensor; among them, the most representative differential sensor is the Dynamic Vision Sensor (DVS). DVS only outputs the pixel address and information whose light intensity changes, so only Sensitive to moving objects, but not to static objects; the integral sensor is currently mainly represented by the retinal sensor (Vidar) led by Professor Huang Tiejun of Peking University, which simulates the cell connection structure and integral distribution in the fovea area of the retina. The principle of pulse is to output the pulses emitted by all pixels at each moment in the form of an array. It has the characteristics of full-time, asynchronous, high-speed, etc., and has sensitive perception of dynamic and static objects; event-based sensors combine DVS with traditional The camera is combined to output a pulsed data stream while also outputting frame images, which can be ATIS (Asynchronous Time-based Image Sensor), DAVIS (Dynamic and Active-pixel Vision Sensor) or CeleX. Therefore, the present invention may also be a target recognition method based on a bionic vision sensor, and may also be a method for rapid target recognition. In the target recognition scheme, the present invention can make the bionic vision sensor play the role of "eyes".

Step 101: Obtain the original pulse data of the photographed target to be identified. The original pulse data includes at least one pulse sequence. In this embodiment, one pulse sequence may include one frame of pulses or multiple frames of pulses, and multiple pulse sequences or multiple frames. The pulse can be used for the subsequent multiple measurement mechanism; as a preferred embodiment, the acquired raw pulse data of the target to be identified comes from a bionic vision sensor, and the bionic vision sensor is used to photograph the target to be identified.

Step 102: Retrieve the area where the target to be identified is located by filtering each pulse in the pulse sequence, and determine the pulse sampling window according to the area where the target to be identified is located, so that the area where the target is located is used as the range of the pulse sampling window. With technology, the present invention only inputs the pulse data in the pulse sampling window into the excitation layer for processing in real time, and does not process the pulses outside the pulse sampling window. Therefore, the present invention can greatly reduce the calculation amount of the inference layer in the pulse neural network. Therefore, the target recognition speed can be significantly accelerated under the same hardware conditions. In some preferred embodiments, after the pulse sampling window is determined, it further includes the step of recording the time interval between two adjacent occurrences of the target to be identified in the pulse sampling window. It should be understood that the time interval here may be more than one time interval. The average value of each time interval or the time interval between two adjacent times at random. For high-speed rotating targets, the present invention can use the above-mentioned fixed sampling window and recording time interval to complete the measurement of the speed of the high-speed rotating target, and then it is the speed of the rotating body. Calibration provides a new way to achieve. The present invention uses a spiking neural network, which relates to the field of brain-like computing. The spiking neural network (SNN) is called the third-generation artificial neural network, and its neuron model is closer to the characteristics of biological neurons. The processing mechanism also borrows more from the brain, using pulses as the medium of information transmission, including time information and spatial information, so the pulse neural network is the most representative algorithm in the field of brain-like computing, and with deep learning With the development of the Spike Neural Network, the connection mode and application scenarios have also changed. For example, in the connection mode, the Spike Neural Network has developed from the early shallow complex connections to the deep simple connections. For example, in the application scenarios, The present invention can apply the spiking neural network to various pattern recognition problems, specifically it can be the recognition of high-speed rotating characters in the present invention. Compared with algorithms such as deep learning and machine learning, the spiking neural network has more biological likelihood. Moreover, relying on its asynchronous response to impulse and processing mechanism, it makes its calculation energy consumption lower and processing speed faster. Spike neuron is the basic component unit and information processing unit of spiking neural network, and it is a kind of biological neuron. A simplified mathematical expression of electrochemical characteristics. The neuron model in the pulsed neural network of the present invention may include an integral excitation model (Integrate and Fire, IF), a leakage current integral excitation model (Leaky Integrate and Fire, LIF), and pulse Response model (Spike Response Model, SRM), etc. In this step, the pulses in the pulse sampling window are input to the pulse neural network. The pulse neural network in the present invention may include an excitation layer and an inference layer.

Step 103, the input pulse is processed by the pulse neural network below. Specifically, this embodiment maps the pulse input to the pulse neural network so that the number of excited neurons in the first layer of the excitation layer corresponds to the size of the pulse sampling window. The mapping method can be hash mapping, hexadecimal conversion, etc. The excitation layer of this embodiment includes multiple layers of excitation neurons connected in sequence through synapses, and the connection between adjacent excitation neurons includes but is not limited to local connections. , Fully connected or dynamically randomly connected, the excitation layer of this embodiment is composed of (Leaky Integrate and Fire) LIF neurons, the first layer of excitation neurons is the input impulse neurons, and the other layers of excitation neurons are the excitation impulse neurons . In specific work, after the excited neuron of the present invention receives a pulse, its membrane potential will change.

Step 104: When the first preset condition is satisfied, the pulse is transmitted along the excitation neurons of each layer except the last layer of excitation neurons. The current excitation neuron is the next layer of excitation neuron or the inference layer emits the pulse, the membrane potential of itself Reset to a resting state. In this embodiment, the first preset condition is that the change in membrane potential of each layer other than the last layer of stimulated neurons exceeds its own first excitation threshold. In this embodiment, the first The excitation threshold is 1.0, so as to achieve noise filtering and pulse enhancement of the pulse input to the pulse neural network; and when the second preset condition is met, the pulse is transmitted to the inference layer along the final layer of the excitation neuron. In this embodiment, the first Second, the preset conditions can be set into two types: 1. The change in membrane potential of the terminal stimulated neuron exceeds its own second excitation threshold; 2. The change in membrane potential of the terminal stimulated neuron exceeds its own second excitation. Threshold and the proportion of the number of activated neurons in the last layer of stimulated neurons (that is, the ratio of the number of stimulated neurons to the number of all neurons in the last layer) is greater than the first preset ratio, the first preset in this embodiment The ratio is 15%. Compared with the first condition, the second condition can adjust the pulse input of the inference layer by controlling the second preset ratio, and the less the pulse input of the inference layer, the amount of calculation required And the lower the power consumption, so the present invention also has the advantage of being able to reduce the calculation amount and power consumption of the pulse processing algorithm. The inference layer includes multiple layers of inference neurons connected in sequence through synapses. In some preferred embodiments of the present invention, please refer to Figures 3 and 4. Figure 3 is a diagram of multiple excitations with local connections in some embodiments of the present invention. A schematic diagram of the working principle of the pulse recognition process using a neuron as an example. FIG. 4 is a schematic diagram of the working principle of the pulse recognition process in some other embodiments of the present invention. Only when the number of activated pulse neurons in the excitation layer exceeds a certain ratio, the emitted pulses are tiled into one dimension and then input to the inference layer uniformly. When there is no target in the sampling window, such as background noise input in the sampling window By adjusting the above-mentioned ratio value, it is possible to realize that when only background noise is input, the excitation layer does not emit pulses to the inference layer, thereby reducing the calculation amount of the entire pulse recognition algorithm and greatly reducing the calculation energy consumption.

The reasoning layer in this embodiment is composed of (Integrate and Fire) IF neurons. Based on the disclosed content of the present invention, the number of layers of the reasoning layer and the number of neurons in each layer can be adjusted reasonably and wisely according to the actual situation. The inference neurons of each layer are connected by synapses, and the synaptic connection methods include but are not limited to convolutional connection, partial connection or full connection. Synaptic strength can be obtained by using existing neural network training methods. Training methods include but It is not limited to gradient-based backpropagation learning algorithms, reinforcement learning algorithms, and pulse time-dependent synaptic plasticity (Spiking Time Dependent Plasticity, STDP) algorithms, etc., and will not be repeated in the present invention. After the inference neuron of the present invention receives the pulse, its membrane potential will change.

Step 105: When the third preset condition is satisfied, the pulse is transmitted along each layer of inference neurons other than the last layer of inference neurons. In this embodiment, the third preset condition is each layer of inference neurons other than the last layer of inference neurons. The change in the membrane potential of the cell itself exceeds its own third excitation threshold, and the third preset value may be 1.0. In the various embodiments provided by the present invention, it should be understood that "sequential transmission" refers to the neuron set in the front transmitting the pulse to the neuron set in the back, as shown in FIG. 3 and FIG. 4.

Step 106: Determine the recognition result according to the activity of the last layer of inference neurons. For example, if the IF neurons in the last layer compare their respective activities to obtain the recognition results, that is, if the activities of the inference neurons in the last layer reach various set thresholds, then the recognition can be considered successful, otherwise, the recognition failed, or, In this embodiment, it is also possible to judge whether the recognition is successful based on whether the sum of the activity of each inference neuron in the final layer reaches a preset value. As some preferred technical solutions, in some embodiments of the present invention, a multiple measurement mechanism can be used, that is, the above steps 101 to 106 can be executed multiple times synchronously or asynchronously, preferably synchronously. It is a pulse sequence output by the simulated vision sensor (the bionic vision sensor breaks the concept of the traditional camera frame and generally has a high time resolution), the original pulse data contains multiple pulse sequences, and the calculation results of the pulse recognition algorithm are not limited One is one, but there are multiple consecutive ones. In this embodiment, the recognition results are finally output only when the multiple recognition results after the above-mentioned processing are performed on all pulse sequences at the same time are consistent. The above-mentioned improved solution can significantly improve the recognition accuracy. Therefore, the target recognition method provided in the foregoing embodiments of the present invention can simultaneously use time information and space information to complete rapid target recognition.

Please refer to FIG. 2, which is a schematic diagram of the composition of a target recognition system in some embodiments of the present invention. The present invention provides a target recognition system. The target recognition system includes a bionic vision sensor and the following target recognition device. The target recognition device can be integrated on a pulse computing platform to implement a pulse processing algorithm, and the bionic vision sensor is used for shooting The target to be identified to obtain the original pulse data (that is, the recorded data) of the target to be identified, and then the recorded data can be output to the pulse calculation platform in real time. When the present invention is implemented, the pulse calculation platform can receive the data output by the bionic vision sensor in real time , The original sensor data needs to be encoded and generally transmitted through the bus, so from the pulse event issuance to the algorithm processing platform to receive, this process produces a small time difference, which can be measured by the current time recording module, which is recorded as Δt1.

The pulse computing platform used in the present invention includes but is not limited to the following types: (1) server, workstation, desktop host or mobile computer, (2) embedded computing platform, such as Multi-Processor System-on-Chip , MPSoC) development board, central processing unit (CPU) development board, graphics processing unit (GPU) development board, single-chip Microcomputer (SCM), etc., (3) field programmable gate Array (Field Programmable Gate Array, FPGA), or Application Specific Integrated Circuit (ASIC), etc., (4) Cloud computing platform, etc. The pulse calculation platform is used to run pulse processing algorithms, which can include system control algorithms and pulse recognition algorithms.

Specifically, some embodiments of the present invention provide a target recognition device, which includes a pulse acquisition module, a sampling window module, a pulse mapping module, a neuron excitation module, a neuron inference module, and a recognition result determination module.

The pulse acquisition module is used to acquire the original pulse data of the target to be identified; the original pulse data contains at least one pulse sequence, where one pulse sequence can include one frame of pulse or multiple frames of pulse, multiple pulse sequences or multiple frames The pulse can be used for subsequent multiple measurement mechanisms.

The sampling window module is used to retrieve the area of the target to be identified by filtering each pulse in the pulse sequence, and is used to determine the pulse sampling window according to the area of the target to be identified, and to sample the pulses in the pulse sampling window. Input to the impulse neural network; the impulse neural network includes an excitation layer and an inference layer. From the algorithm processing platform receiving the pulse data, to the sampling window module filtering out the pulse data in the sampling window and inputting it into the excitation layer, this process produces a slight time difference, which can be measured by the current moment recording module and recorded as Δt2.

The pulse mapping module is used to map the pulses input to the pulse neural network so that the number of excited neurons in the first layer of the excitation layer corresponds to the size of the pulse sampling window. The pulse mapping module can be understood as the pulse recognition algorithm used in the present invention The impulse event mapping layer of the impulse neural network can use hash mapping, binary conversion and other methods to map impulse events to the input impulse neurons of the excitation layer; the excitation layer contains multiple layers of excitation nerves connected in sequence through synapses Meta, the connection mode of the synapse can include partial connection, full connection or dynamic random connection, etc. The following content takes the local connection mode as an example for specific description.

First of all, the number of input pulse neurons in the above excitation layer needs to be consistent with the size of the pulse sampling window. Assuming that the sampling window size is N*N, the input pulse in the sampling window is actually corresponding to the position of N*N bionic vision sensors. The photoreceptor neuron is excited, and the input pulse neuron that reaches the excitation layer has a time delay of (Δt1+Δt2). As shown in Figure 3, each input impulse neuron is locally connected to multiple excitation impulse neurons. The local connection mode is: the input impulse neuron's own position as the center, the pulse in the surrounding neighborhood with R as the radius Neurons are connected (if the input neuron is located at the boundary of the sampling window, only the impulse neurons within the range of N*N are connected).

The neuron excitation module is used to enable the pulse to be transmitted along the firing neurons in each layer other than the terminal layer when the first preset condition is met, and to enable the pulse to be transmitted along the terminal layer to stimulate the neuron when the second preset condition is met To the inference layer; the inference layer contains multiple layers of inference neurons connected in sequence. Specifically, the first preset condition is that the change in membrane potential of each layer other than the final layer of stimulated neurons exceeds its own first excitation threshold, and the second preset condition is that the membrane potential of the final layer of stimulated neurons itself The amount of change exceeds its second excitation threshold and the proportion of the number of activated neurons in the final layer of excitation neurons is greater than the first preset ratio. Specifically, when the input impulse neuron emits a pulse, the membrane potential of the excitation impulse neuron connected to it and located at the same position increases by Δ1 (that is, the connected synapse strength is Δ1), and the other excitation impulse neurons connected to it have an increase in membrane potential. The membrane potential increases by Δ2 (that is, the connected synaptic strength is Δ2), where Δ1>Δ2, when there is no pulse input, the membrane potential of the stimulated impulse neuron connected to it and located at the same position attenuates Δ3 (that is, the membrane of the LIF neuron The potential attenuation factor is Δ3); among them, the values of Δ1, Δ2, and Δ3 can be preset as fixed values or can be dynamically adjusted. When the membrane potential of a stimulated neuron exceeds its own excitation threshold, the neuron activates and emits a pulse, and at the same time resets its membrane potential to a resting state.

The neuron inference module is used to make the pulses be transmitted sequentially along each layer of inference neurons except the last layer of inference neurons when the third preset condition is satisfied. Specifically, the third preset condition is that the change in membrane potential of each layer of inference neurons other than the last layer of inference neurons exceeds its own third excitation threshold. Specifically, unlike the LIF neurons in the excitation layer, the IF neurons in the inference layer do not have the characteristics of membrane potential attenuation. After the IF neurons in the inference layer receive a pulse input, their own membrane potential increases, and the value of the increase is equal to the total pulse input If the membrane potential of the IF neuron exceeds its own excitation threshold, a pulse will be emitted and transmitted to the next layer of IF neurons connected to it. At the same time, the membrane potential of the IF neuron will be reset, and the pulse will go from layer to layer. After spreading, until the last layer.

The recognition result determination module is used to determine the recognition result according to the activity of the last layer of inference neurons. For example, if the IF neurons in the last layer compare their respective activities to obtain the recognition results, that is, if the activities of the inference neurons in the last layer reach various set thresholds, then the recognition can be considered successful, otherwise, the recognition failed, or, In this embodiment, it is also possible to judge whether the recognition is successful based on whether the sum of the activity of each inference neuron in the final layer reaches a preset value.

In some improved embodiments of the present invention, the target identification device may further include a rotational speed measurement module and a current moment recording module. The rotational speed measurement module is used to record that the target to be identified appears twice in the pulse sampling window after the pulse sampling window is determined. The rotational speed measurement module fixes the sampling window at a position within the camera's field of view according to the rotation radius of the rotating target, and then accumulates the time interval during which the same target appears in the sampling window multiple times, and can reduce the error by statistical methods , Calculate the time required for the target to rotate one circle, and then realize the measurement of the target speed of high-speed rotation. The current time recording module is used to accurately obtain the current time. The specific form of the time can be selected reasonably and wisely according to needs. The present invention can make the delay of time accuracy as small as possible. The invention adopts the pulse neural network to process the pulse sequence output by the bionic vision sensor, and the processing method simultaneously utilizes the time information and the space information, can realize the rapid identification of the target, and is still applicable under the condition of high movement speed.

In some embodiments of the present invention, a computer-readable storage medium may also be provided, and a computer program is stored on the computer-readable storage medium, and the computer program may be executed by a processor, so as to realize the The target recognition method, the present invention can make the above-mentioned computer program run on the impulse computing platform.

In the following, the working process of the present invention and the significant effects actually brought by it will be further described in detail with the specific experimental process. This embodiment uses the above-mentioned integral bionic vision sensor Vidar, the sensor has a time resolution of 25 microseconds, and a display resolution of 400*250 pixels. The sensor is used to shoot at a fine noon on a high-speed rotating industrial fan. The center distance between the fan and the sensor is about 60cm, and the fan radius is about 20cm. There are 3 white English characters on black background ('P'). ,'C','L'), the blade rotation speed is about 2500R/min, the captured pulse data is packaged on the data acquisition card and then transmitted to the desktop computer through the USB3.0 bus for processing, and the desktop computer receives and obtains Vidar in real time The output data, the sampling window setting module filters the pulse data, and the pulse event mapping layer maps the filtered pulse data to the input pulse neurons of the excitation layer one by one. The position of the sampling window is set according to the position of the English character in the Vidar field of view The size of the sampling window depends on the pixel size of the English characters in Vidar's field of view. In this embodiment, the coordinates (y, x) of the upper left corner of the sampling window are (50, 180), and the size of the sampling window is 40*40 pixels. That is, the excitation layer contains 40*40 input pulse neurons. The pulse processing algorithm is effective for bionic vision. The pulse data input by the sensor is processed. As shown in Figure 4, the number of pulse input neurons in the excitation layer is the same as the sampling window size, which is 40*40. The pulse event mapping module realizes the pulse input in the sampling window and the pulse input neuron One-to-one mapping. Each impulse input neuron is connected to 9 excitation impulse neurons by synapses. The synapses are connected as follows: excitation in the surrounding eight neighborhoods centered on the input impulse neuron's own position The impulse neurons are connected (if the input impulse neuron is located at the boundary of the sampling window, only the impulse neurons within the range of 40*40 are connected). When the input impulse neuron emits a pulse, the membrane potential of the excitation impulse neuron connected to it and located at the same position increases by 1.4 (Δ1), and the membrane potential of the other excitation impulse neurons connected to it increases by 0.5 (Δ2). When there is no pulse input, the membrane potential of the stimulated pulse neuron connected to it and located at the same position attenuates by 0.25 (Δ3). When the membrane potential of the excitation pulse neuron exceeds its own excitation threshold (1.0), the excitation pulse neuron activates and emits a pulse. Please refer to Figures 5 to 7, which uses the excitation layer provided by the present invention to perform noise on the pulse sequence. The effect of filtering and pulse enhancement is shown. The picture on the left is the pulse input of the bionic vision sensor, and the picture on the right is the effect of filtering and enhancement using the excitation layer of the present invention. Specifically, the white part represents the position of the pulse input and the black part It represents the position where there is no pulse input, and the box area represents the pulse sampling window; among them, Figure 5 is the effect of filtering and enhancing the pulse sequence of the region where the character "P" is located through the excitation layer. Figure 6 shows the effect of the pulse sequence in the region where the character "P" is located through the excitation layer. The effect of noise filtering and pulse enhancement on the pulse sequence of the area where the target to be recognized—the character "C" is performed by the layer. Figure 7 shows the pulse sequence of the area where the target to be recognized—the character "L" is passed through the excitation layer for noise filtering and pulse The enhanced effect is that only when the number of activated excitation pulse neurons exceeds a certain proportion (15%), the excitation pulses are tiled into one dimension (1*1600) and then uniformly input to the inference layer. When the rotating character does not appear in the sampling window, the input in the sampling window is background noise; as shown in Figure 4, the inference layer is composed of 3 layers of IF neurons, and the number of IF neurons in each layer is 512, 1024, and 3. The IF neurons between each layer realize synaptic connection through a fully connected way. After the IF neuron receives the pulse input, its own membrane potential increases, and the increased value is equal to the sum of the synaptic strength values of all the pulse inputs. If the membrane potential of an IF neuron exceeds its own excitation threshold (1.0), a pulse is sent and transmitted to the next layer of IF neurons connected to it, and its membrane potential is reset to a resting state (0). The pulse propagates back layer by layer until it reaches the last layer. The IF neuron of the last layer obtains the recognition result by comparing the activity of each neuron. The multiple measurement mechanism of the present invention, for example, only when the recognition results outputted by the inference layer for 5 consecutive times are consistent, the final recognition result can be obtained. This method can significantly improve the recognition accuracy and recognition speed, and realize the measurement of the rotation speed of high-speed rotating characters. . The invention significantly improves the recognition speed through the optimization method of multi-threaded parallel computing, splits the computing task to multiple processors of the desktop computer for simultaneous calculation, and records the time interval when the same character appears in the sampling window twice in succession. Calculate the time required for the character to rotate one circle, and then realize the measurement of the speed of the high-speed rotating fan.

In the description of this specification, the description with reference to the terms "this embodiment", "one embodiment", "some embodiments", "examples", "specific examples", or "some examples", etc. means to combine the embodiments The specific features, structures, materials or characteristics described by the examples are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the described specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine the different embodiments or examples and the features of the different embodiments or examples described in this specification without contradicting each other.

In addition, the terms "first" and "second" are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined with "first" and "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, "plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.

The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modification, equivalent replacement and simple improvement made in the essence of the present invention should be included in the protection scope of the present invention. Inside.

Claims (10)

1. A target recognition method, characterized in that: the method includes the following steps;

   Acquiring raw pulse data of the photographed target to be identified, where the raw pulse data includes at least one pulse sequence;

   The area where the target to be identified is located is retrieved by screening each pulse in the pulse sequence, the pulse sampling window is determined according to the area where the target to be identified is located, and the pulses in the pulse sampling window are input to the pulse neural network, the pulse neural network Including excitation layer and reasoning layer;

   Mapping the pulses input to the spiking neural network so that the number of firing neurons in the first layer of the firing layer corresponds to the size of the pulse sampling window, and the firing layer includes multiple layers of firing neurons connected in sequence;

   When the first preset condition is met, the pulse is transmitted along the firing neurons of each layer other than the final layer in turn, and the pulse is transmitted to the inference layer along the firing neuron of the last layer when the second preset condition is met, the reasoning layer Contains multiple layers of inference neurons connected in sequence;

   When the third preset condition is satisfied, the pulse is transmitted in sequence along each layer of inference neurons except the last layer of inference neurons;

   The recognition result is determined according to the activity of the last layer of inference neurons.
2. The target recognition method according to claim 1, wherein the acquired raw pulse data of the target to be recognized comes from a bionic vision sensor, and the bionic vision sensor is used to photograph the target to be recognized.
3. The target recognition method according to claim 1 or 2, characterized in that: the first preset condition is that the change in membrane potential of each layer of stimulated neurons other than the final layer of stimulated neurons exceeds its own first excitation threshold .
4. The target recognition method according to claim 3, characterized in that: the second preset condition is that the change in membrane potential of the final layer of stimulated neurons exceeds its own second excitation threshold and the final layer of stimulated neurons is activated The proportion of the number of neurons is greater than the first preset proportion.
5. The target recognition method according to any one of claims 1, 2 or 4, wherein the third preset condition is the change in membrane potential of each layer of inference neurons other than the last layer of inference neurons. Exceeds its own third excitation threshold.
6. The target recognition method according to claim 1, characterized in that: after the pulse sampling window is determined, it further comprises the step of recording the time interval when the target to be recognized appears in the pulse sampling window twice.
7. The target recognition method according to any one of claims 1, 2, 4, or 6, wherein the original pulse data contains a plurality of pulse sequences, and all the pulse sequences are simultaneously identified after the above processing The recognition result is output only when the results are consistent.
8. A target recognition device, characterized in that: the device includes a pulse acquisition module, a sampling window module, a pulse mapping module, a neuron excitation module, a neuron inference module, and a recognition result determination module;

   The pulse acquisition module is used to acquire the original pulse data of the target to be identified that is photographed; the original pulse data includes at least one pulse sequence;

   The sampling window module is used to retrieve the area where the target to be recognized is located by filtering each pulse in the pulse sequence, and is used to determine the pulse sampling window according to the area where the target to be recognized is located, and to sample the pulse in the window The pulses of are input to the impulse neural network; the impulse neural network includes an excitation layer and an inference layer;

   The pulse mapping module is used to map the pulses input to the pulse neural network so that the number of the first layer of excitation neurons in the excitation layer corresponds to the size of the pulse sampling window; the excitation layer contains multiple layers of sequentially connected excitation nerves Yuan;

   The neuron excitation module is used to enable the pulse to be transmitted along the firing neurons in each layer other than the terminal layer when the first preset condition is satisfied, and to enable the pulse to fire the nerve along the terminal layer when the second preset condition is satisfied Elements are transferred to the inference layer; the inference layer includes multiple layers of inference neurons connected in sequence;

   The neuron inference module is used to enable pulses to be sequentially transmitted along each layer of inference neurons other than the last layer of inference neurons when the third preset condition is satisfied;

   The recognition result determination module is used to determine the recognition result according to the activity of the last layer of inference neurons.
9. A target recognition system, characterized in that: the target recognition system comprises a bionic vision sensor and the target recognition device according to claim 8, the bionic vision sensor is used to photograph the target to be recognized to obtain the original pulse data of the target to be recognized .
10. A computer-readable storage medium with a computer program stored thereon, characterized in that the computer program is executed by a processor to implement the target identification method according to any one of claims 1-7.

Images