Image classification method and system based on hierarchical competitive spiking neural network

Abstract

The invention relates to an image classification method and system based on a hierarchical competitive spiking neural network, belongs to the field of image classification, solves the problem of low recognition accuracy during image classification, and comprises the following steps: inputting preprocessed image data into a trained SNN network to obtain a prediction classification label of an image; wherein a competition strategy is introduced into the SNN, and the competition strategy comprises the following steps: sorting feature maps of different positions of input image data, and realizing excitation or suppression of the feature maps; and sorting different spiking neurons based on the sensitivity to realize excitation or suppression of the spiking neurons. The training of the SNN network comprises the following steps of: allocating a weight to each convolution kernel in the network, and allocating a weight to each neuron; calculating to obtain a training loss value; and carrying out back propagation, and calculating to obtain the weight of each convolution kernel, the weight of each neuron, an excitation coefficient and a suppression coefficient. The feature expression ability of the pulse neural network is enhanced, and the accuracy of image recognition is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image classification, and in particular to an image classification method and system based on a layered competitive pulse neural network. Background Art

[0002] Traditional image classification methods rely on hand-crafted features, including the color and texture of the image, which are difficult to fully capture the complexity and diversity of the image. In addition, traditional image classification methods often lack robustness and generalization capabilities, and the classification results may be different for images in different scenes and lighting conditions.

[0003] The Spiking Neural Network (SNN) is a third-generation neural network computing model. It uses discrete pulses instead of traditional continuous activation values ​​to encode and transmit information. It is a neural network model that is closer to the biological neural processing mechanism and has higher computational efficiency. When using traditional SNN networks for image classification tasks, there are two main problems. First, the information transmission between SNN network layers uses discrete pulses instead of traditional continuous activation values ​​to encode and transmit information, resulting in weak feature expression capabilities and low accuracy of image classification results. Second, the use of a spiking neuron model instead of the original activation function cannot be trained using the traditional gradient descent method. At the same time, the concept of time steps is introduced, which makes training difficult and affects the efficiency of image classification. Summary of the invention

[0004] In view of the above analysis, an embodiment of the present invention aims to provide an image classification method and system based on a hierarchical competitive pulse neural network, so as to solve the problems of inaccurate classification results of existing image classification methods, weak feature expression ability of existing SNN networks in image classification, and high training difficulty.

[0005] On the one hand, an embodiment of the present invention provides an image classification method based on a layered competitive spiking neural network, comprising:

[0006] Preprocess the image data;

[0007] The preprocessed image data is input into the trained SNN network to obtain the predicted classification label of the image; wherein a competition strategy is introduced into the SNN network, including:

[0008] Passing the input image data through a convolution layer to obtain a feature map, sorting the feature response of each position in the feature map, and using the sorting result to achieve excitation or inhibition of the feature map, wherein the magnitude of the excitation or inhibition is obtained by the excitation coefficient and the inhibition coefficient respectively; and sorting different pulse neurons based on sensitivity, and using the sorting result to achieve excitation or inhibition of the pulse neurons;

[0009] The training of the SNN network includes: assigning weights to each convolution kernel in the convolution layer of the SNN network and assigning weights to each neuron in the fully connected layer; calculating the loss value of the SNN network training; using the loss value for back propagation to calculate the weights of each convolution kernel, the weights of each neuron, the excitation coefficient and the inhibition coefficient.

[0010] Based on the further improvement of the above method, the feature maps at different positions of the input image data are sorted, and the sorting result is used to realize the stimulation or suppression of the feature maps, including:

[0011] The input image data is passed through a convolution layer to obtain a feature map, and the feature responses at each position in the feature map are divided into three categories according to the response strength, wherein: using the first mapping function, the k1 features with the strongest responses are given positive excitation based on the excitation coefficient, the features whose responses are ranked after k1 and whose response values ​​are above 0 are suppressed based on the suppression coefficient, and the features whose responses are less than 0 are set to 0.

[0012] Based on the further improvement of the above method, the following formula is used to implement the sorting of the feature maps at different positions of the input image data, and the sorting result is used to implement the stimulation or suppression of the feature map:

[0013]

[0014] Among them, t represents the time step in the SNN network, n represents the nth feature map obtained after the feature extraction by the convolution layer, and the first mapping function represents the value after the feature value at the (x, y) position in the nth feature map is mapped, represents the feature response at the (x, y) position in the nth feature map, represents the value of the k1th feature response ranking of the (x, y) position in all feature maps obtained through the convolution layer, wherein the ranking is arranged in descending order of feature values; β r and β l are hyperparameters, representing the excitation coefficient and the inhibition coefficient respectively.

[0015] A further improvement based on the above method, wherein different spiking neurons are sorted based on sensitivity, and the spiking neurons are stimulated or inhibited using the sorting results, including:

[0016] Using the second mapping function, the k2 spiking neurons that are most sensitive to the input information retain their original responses, the spiking neurons whose sensitivity to the input information is ranked after k2 and whose sensitivity is above 0 are inhibited from responding, and the spiking neurons that have no response to the input information are set to 0.

[0017] Based on the further improvement of the above method, the following formula is used to implement the sorting of different spiking neurons based on sensitivity, and the sorting result is used to implement the stimulation or inhibition of the spiking neurons:

[0018]

[0019] Among them, t represents the time step in the SNN network, and the second mapping function o t,m represents the response value output by the mth pulse neuron after mapping, f t,m represents the output response value of the mth spiking neuron, f t,to It represents the k2th response value obtained by sorting the response values ​​of each pulse neuron in the fully connected layer from large to small.

[0020] Based on a further improvement of the above method, the step of assigning weights to each neuron in the fully connected layer includes: during the back propagation process, calculating the global learning weights, local learning weights, and the weight allocation ratio between the global learning weights and the local learning weights of the neurons.

[0021] Based on the further improvement of the above method, the weights of each neuron in the fully connected layer are calculated using the following formula:

[0022]

[0023] Among them, ∝ is a hyperparameter, which represents the weight distribution ratio, represents the global learning weight of the corresponding neuron, represents the local learning weight of the corresponding neuron, l represents the number of spiking neurons in the fully connected layer, and N represents the number of convolutional layers.

[0024] Based on the further improvement of the above method, the training of the SNN network further includes:

[0025] The STDP rule is adopted to utilize the pulse correlation of the pulse neurons in the front and back layers to enhance or weaken the synaptic weight, including: when the pulse correlation of the pulse neurons in the front and back layers has a temporal causal relationship, the synaptic connection is enhanced; otherwise, the synaptic connection is weakened.

[0026] On the other hand, an embodiment of the present invention provides an image classification system based on a hierarchical competitive spiking neural network, including an image acquisition module, an image preprocessing module, an SNN network module and a training module, wherein:

[0027] Image acquisition module: used to acquire images;

[0028] Image preprocessing module: used to preprocess the collected image data;

[0029] SNN network module: used to calculate the predicted classification label of the image using the preprocessed image data; the SNN network module introduces a competition strategy, including:

[0030] Passing the input image data through a convolution layer to obtain a feature map, sorting the feature response of each position in the feature map, and using the sorting result to achieve excitation or inhibition of the feature map, wherein the magnitude of the excitation or inhibition is obtained by the excitation coefficient and the inhibition coefficient respectively; and sorting different pulse neurons based on sensitivity, and using the sorting result to achieve excitation or inhibition of the pulse neurons;

[0031] Training module: used to assign weights to each convolution kernel in the convolution layer of the SNN network and to each neuron in the fully connected layer; calculate the loss value of the SNN network training; use the loss value for back propagation to calculate the weights of each convolution kernel, the weights of each neuron, the excitation coefficient and the inhibition coefficient.

[0032] Based on the further improvement of the above system, the weights of each neuron in the fully connected layer are calculated using the following formula:

[0033]

[0034] Among them, ∝ is a hyperparameter, which represents the weight distribution ratio, represents the global learning weight of the corresponding neuron, represents the local learning weight of the corresponding neuron, l represents the number of spiking neurons in the fully connected layer, and N represents the number of convolutional layers.

[0035] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0036] 1. The present invention adopts an image classification method based on a hierarchical competitive pulse neural network. By introducing a hierarchical heterogeneous competition strategy into the traditional SNN network, different competition mechanisms are designed for the convolutional layer and the fully connected layer. The convolutional layer uses the first mapping function to learn local features, and the fully connected layer uses the second mapping function to aggregate global features, thereby enhancing the SNN network's ability to express features of image data, allowing the network to learn richer and more diverse feature representations, capturing the complexity and diversity of images, and improving the model's ability to distinguish and the accuracy of image classification.

[0037] 2. The present invention proposes an adaptive joint learning method. The method organically combines a global back propagation algorithm based on supervised learning, a local synaptic plasticity algorithm based on unsupervised learning, and a local unsupervised learning algorithm based on competitive learning. Local learning rules are used during network forward propagation; global learning rules are used to adjust parameters during backward propagation. In addition, the method can also adaptively adjust the weight distribution of global learning and local learning, reduce the manual parameter adjustment process, and make network training more robust to tasks and data sets.

[0038] 3. The adaptive joint learning algorithm introduced in the present invention can use the global feedback signal to quickly provide training direction in the early stage and improve the convergence speed; and then increase the dependence on local learning in the later stage. Compared with a single learning rule, it can obtain a better training curve, significantly reduce redundant calculations, and improve the computational efficiency of image recognition.

[0039] In the present invention, the above-mentioned technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can become obvious from the description, or can be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. In the entire drawings, the same reference symbols represent the same components;

[0041] Figure 1 A flowchart of an image classification method based on a layered competitive spiking neural network is shown in one embodiment of the present invention;

[0042] Figure 2 A structural diagram of a WRLP competition strategy shown in an embodiment of the present invention;

[0043] Figure 3 A structural diagram of a KWTA competition strategy shown in one embodiment of the present invention;

[0044] Figure 4 A flowchart of an adaptive joint learning method according to an embodiment of the present invention;

[0045] Figure 5 A block diagram of an image classification system based on a layered competitive spiking neural network is shown as an embodiment of the present invention. DETAILED DESCRIPTION

[0046] The preferred embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0047] A specific embodiment of the present invention discloses an image classification method based on a layered competitive spiking neural network. Figure 1 As shown, including:

[0048] S11: Preprocess the image data.

[0049] Here, image refers to a picture taken by a camera or a picture downloaded from a website;

[0050] Specifically, the collected image data is normalized, denoised and cropped to ensure the quality and consistency of the input image data. S12: The preprocessed image data is input into the trained SNN network to obtain the predicted classification label of the image; the image classification label exemplarily adopts common single labels, including people, animals, plants, food, transportation, and daily objects, and the predicted classification label of the image is obtained using the SNN network.

[0051] Wherein, introducing a competition strategy into the SNN network includes:

[0052] Passing the input image data through a convolution layer to obtain a feature map, sorting the feature response of each position in the feature map, and using the sorting result to achieve excitation or inhibition of the feature map, wherein the magnitude of the excitation or inhibition is obtained by the excitation coefficient and the inhibition coefficient respectively; and sorting different pulse neurons based on sensitivity, and using the sorting result to achieve excitation or inhibition of the pulse neurons;

[0053] The training of the SNN network includes: assigning weights to each convolution kernel in the convolution layer of the SNN network and assigning weights to each neuron in the fully connected layer; calculating the loss value of the SNN network training; using the loss value for back propagation to calculate the weights of each convolution kernel, the weights of each neuron, the excitation coefficient and the inhibition coefficient.

[0054] It should be noted that, based on the characteristics of different network layers in the SNN network, this patent proposes a hierarchical heterogeneous competition strategy, which adopts different competition strategies for different structural levels of the SNN network. When a neuron generates a pulse, it will inhibit the activity of other surrounding neurons, so that different neurons gradually specialize to respond to different input patterns, thereby achieving the purpose of enhancing the network's expression ability. This embodiment designs different competition mechanisms for the convolutional layer and the fully connected layer to match their respective computing characteristics. Under this competition strategy, the SNN network produces feature differentiation and sparse coding under unsupervised conditions, simulating the development process of the biological nervous system.

[0055] Specifically, when the input preprocessed image data passes through the convolution layer, each convolution kernel in the convolution layer extracts local features from the image data to obtain multiple sets of feature maps. Since the convolution layer learns local features, the competitive learning process must also be localized to enhance the learning of local features. To achieve this, the present invention sorts the feature maps at different positions of the input image data, and uses the sorting results to achieve incentives or inhibition of the feature maps, which is called the Winner Reward Loser Punish (WRLP) competition strategy. Figure 2 As shown, it includes: passing the input image data through a convolution layer to obtain a feature map, and dividing the feature response of each position in the feature map into three categories according to the response strength, wherein: using the first mapping function, the k1 features with the strongest response are given positive excitation based on the excitation coefficient, and the features whose response is ranked after k1 and whose response value is above 0 are suppressed based on the suppression coefficient, and the features whose response is less than 0 are set to 0. Through this mapping method, the feature values ​​of the convolution layer can be discretized and learned and expressed in the feature space, reducing information loss and improving feature learning ability.

[0056] The WRLP competition strategy for a convolutional layer in the network is implemented using the following formula:

[0057]

[0058] Among them, t represents the time step in the SNN network, n represents the nth feature map obtained after the feature extraction by the convolution layer, and the first mapping function represents the value after the feature value at the (x, y) position in the nth feature map is mapped, represents the feature response at the (x, y) position in the nth feature map, represents the value of the k1th feature response ranking of the (x, y) position in all feature maps obtained through the convolution layer, wherein the ranking is arranged in descending order of feature values; β r and β l are hyperparameters, representing the excitation coefficient and the inhibition coefficient respectively; when β r and β l When both are 0, it means that no competition strategy is adopted. Input the LIF neuron of the SNN network. The value of K1 can be obtained through multiple experimental comparisons. For example, K1 is 5.

[0059] The present invention adopts the WRLP competition strategy for each convolution layer in the SNN network. The data output by the last convolution layer passes through the pooling layer, flattening layer, and linear layer in sequence and then input into the fully connected layer for feature aggregation. In the fully connected layer, the present invention adopts the "k Winners Take All" (kWTA) strategy to perform competitive learning of global features to enhance its performance. Since the fully connected layer has a global receptive field, the present invention adopts a more rigorous method to allow competition between neurons in the layer.

[0060] Specifically, the method of sorting different spiking neurons based on sensitivity and using the sorting results to stimulate or inhibit spiking neurons is called the kWTA competition strategy, such as Figure 3 As shown, it includes: using the second mapping function, retaining the original response of the k2 pulse neurons that are most sensitive to the input information, inhibiting the response of the pulse neurons whose sensitivity to the input information is ranked after k2 and whose sensitivity is above 0, and setting the pulse neurons that do not respond to the input information to 0; wherein sensitivity refers to the degree of response of the output of the pulse neuron to the change of the input pulse signal. The value of K2 can be obtained by multiple experimental comparisons, and K2 is 5 for example.

[0061] The response value of the fully connected layer using the kWTA strategy is obtained by the following formula:

[0062]

[0063] Among them, t represents the time step in the SNN network, and the second mapping function o t,m represents the response value output by the mth pulse neuron after mapping, f t,m represents the output response value of the mth spiking neuron, f t,topk2 represents the k2th response value obtained by sorting the response values ​​of each pulse neuron in the fully connected layer from large to small. The kWTA strategy is more stringent than the WRLP strategy and is conducive to improving the accuracy of the classification task. t,m Input the LIF neuron of the SNN network and finally output the predicted classification label of the image.

[0064] In the training process of the SNN network, the present invention adopts an adaptive joint learning method, combining local unsupervised learning, global supervised learning and competitive learning, such as Figure 4 As shown. The preprocessed image data and the corresponding labels are used as training sets and input into the SNN network. By comparing the predicted classification labels of the images output by the SNN network with the labels actually corresponding to the images, the loss value of the SNN network training is obtained, the network parameters are adjusted according to the global back propagation algorithm and the local synaptic plasticity algorithm, and the learning weights are adaptively adjusted through the weight allocation ratio, including:

[0065] In the forward propagation process of the SNN network, the local learning rules and competition strategies are used to adjust the synaptic weights according to the temporal correlation of neuron pulses; illustratively, the local learning rules adopt the STDP (Spike Timing Dependent Plasticity) rule, and the pulse correlation of the front and rear layer pulse neurons is used to enhance or weaken the synaptic weights, including: when the pulse correlation of the front and rear layer pulse neurons has a temporal causal relationship, the synaptic connection is strengthened; otherwise, the synaptic connection is weakened;

[0066] In the backward propagation process of the SNN network, the event-based back propagation algorithm BPTT is used, which propagates the error signal only when a pulse occurs, minimizing the output classification error through gradient descent. Back propagation adjusts the forward connection weights between layers to achieve better pattern recognition. The weight distribution ratio autonomously adjusts the fusion of global and local learning rules according to different data sets, improving the model's adaptability to different task scenarios.

[0067] In addition, this embodiment incorporates competitive parameters and the weight distribution ratio between global and local learning methods into the back-propagation update, which enables the network to adaptively and collaboratively perform a given pattern recognition task, thereby achieving adaptive control of the entire network.

[0068] Specifically, the preprocessed image X t Input SNN network, t represents the time step in SNN network; the present invention sets a global learning weight for each convolution kernel in the convolution layer Where n represents the convolution layer number, n ranges from 1 to N, N represents the number of convolution layers, and c represents the serial number of the convolution kernel in the corresponding convolution layer; the extracted feature data is obtained through the convolution layer of the SNN network

[0069] Using WRLP competition strategy to obtain Mapping, we get

[0070] Sequentially pass through the N convolutional layers of the SNN network, and use the WRLP competition strategy for each convolutional layer;

[0071] The feature data output by the last convolutional layer Mapping After passing through the pooling layer, flattening layer, and linear layer, the data is input into the fully connected layer for feature aggregation, including:

[0072] Set a learning weight W for each neuron in the fully connected layer N+1 ,right Weighted Wherein, l represents the number of spiking neurons in the fully connected layer; the learning weight includes the global learning weight and the local learning weight of the neuron;

[0073] Using kWTA competition strategy to Mapping, we get

[0074] Compare the predicted classification label of the image output by the SNN network with the label corresponding to the image, and get the loss value L of the SNN network training t .

[0075] Using L t Perform back propagation and calculate W N+1 , β l , β r and ∝.

[0076] Furthermore, weights W are assigned to each neuron in the fully connected layer. N+1 ,include:

[0077] During the back-propagation process, the global learning weights, local learning weights, and weight distribution ratio between the global learning weights and the local learning weights of the neurons are calculated.

[0078] The weights of each neuron in the fully connected layer are calculated using the following formula:

[0079]

[0080] Among them, ∝ is a hyperparameter, which represents the weight distribution ratio, represents the global learning weight of the corresponding neuron, represents the local learning weight of the corresponding neuron, l represents the number of spiking neurons in the fully connected layer, and N represents the number of convolutional layers; ∝, W N+1 , and Learned using back-propagation.

[0081] Compared with the prior art, this embodiment provides an image classification method based on a hierarchical competitive spiking neural network. By introducing a hierarchical heterogeneous competition strategy into the traditional SNN network, different competition mechanisms are designed for the convolution layer and the fully connected layer. The convolution layer uses a first mapping function to learn local features; the fully connected layer uses a second mapping function to aggregate global features, thereby enhancing the SNN network's ability to express features of image data, allowing the network to learn richer and more diverse feature representations, capturing the complexity and diversity of images, and improving the model's ability to distinguish images. An adaptive joint learning method is used to combine the monitoring-based The global back propagation algorithm of supervised learning, the local synaptic plasticity algorithm based on unsupervised learning, and the local unsupervised learning algorithm based on competitive learning are organically combined. That is, local learning rules are used in forward network propagation, and global learning rules are used to adjust parameters in backward propagation, so as to realize adaptive adjustment of weight distribution of global learning and local learning, reduce manual parameter adjustment process, and make network training more robust to tasks and data sets; in the early stage, global feedback signals are used to quickly provide training direction and improve convergence speed; in the later stage, the dependence on local learning is increased, which can obtain better training curve, significantly reduce redundant calculation and improve the computational efficiency of image recognition.

[0082] A specific embodiment of the present invention discloses an image classification system based on a SNN network, such as Figure 5 As shown, it includes: an image acquisition module 51, an image preprocessing module 52, an SNN network module 53 and a training module 54, wherein:

[0083] Image acquisition module 51: used for acquiring images;

[0084] Image preprocessing module 52: used to preprocess the collected image data;

[0085] SNN network module 53: used to calculate the predicted classification label of the image using the preprocessed image data; the SNN network module introduces a competition strategy, including:

[0086] Passing the input image data through a convolution layer to obtain a feature map, sorting the feature response of each position in the feature map, and using the sorting result to achieve excitation or inhibition of the feature map, wherein the magnitude of the excitation or inhibition is obtained by the excitation coefficient and the inhibition coefficient respectively; and sorting different pulse neurons based on sensitivity, and using the sorting result to achieve excitation or inhibition of the pulse neurons;

[0087] Training module 54: used to assign weights to each convolution kernel in the convolution layer of the SNN network and to each neuron in the fully connected layer; calculate the loss value of the SNN network training; use the loss value to perform back propagation to calculate the weight of each convolution kernel, the weight of each neuron, the excitation coefficient and the inhibition coefficient.

[0088] It should be noted that the system and method of the present invention are based on the same inventive concept, and the specific contents can be referenced from each other and will not be repeated here; including network training process, preprocessing process, etc.

[0089] Those skilled in the art will appreciate that all or part of the processes of the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, wherein the computer-readable storage medium is a disk, an optical disk, a read-only storage memory, or a random access memory, etc.

[0090] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. An image classification method based on a hierarchical competitive spiking neural network, characterized in that: include: Preprocess the image data; The preprocessed image data is input into the trained SNN network to obtain the predicted classification label of the image; wherein a competition strategy is introduced into the SNN network, including: Passing the input image data through a convolution layer to obtain a feature map, sorting the feature response of each position in the feature map, and using the sorting result to achieve excitation or inhibition of the feature map, wherein the magnitude of the excitation or inhibition is obtained by the excitation coefficient and the inhibition coefficient respectively; and sorting different pulse neurons based on sensitivity, and using the sorting result to achieve excitation or inhibition of the pulse neurons; The training of the SNN network includes: assigning weights to each convolution kernel in the convolution layer of the SNN network and assigning weights to each neuron in the fully connected layer; calculating the loss value of the SNN network training; using the loss value for back propagation to calculate the weights of each convolution kernel, the weights of each neuron, the excitation coefficient and the inhibition coefficient.

2. The image classification method based on hierarchical competitive pulse neural network according to claim 1 is characterized in that: The step of sorting the feature maps at different positions of the input image data and using the sorting result to stimulate or inhibit the feature maps includes: The input image data is passed through a convolution layer to obtain a feature map, and the feature responses at each position in the feature map are divided into three categories according to the response strength, wherein: using the first mapping function, the k1 features with the strongest responses are given positive excitation based on the excitation coefficient, the features whose responses are ranked after k1 and whose response values ​​are above 0 are suppressed based on the suppression coefficient, and the features whose responses are less than 0 are set to 0.

3. The image classification method based on hierarchical competitive pulse neural network according to claim 2 is characterized in that: The following formula is used to sort the feature maps at different positions of the input image data, and the sorting result is used to stimulate or inhibit the feature map: Among them, t represents the time step in the SNN network, n represents the nth feature map obtained after the feature extraction by the convolution layer, and the first mapping function represents the value after the feature value at the (x, y) position in the nth feature map is mapped, represents the feature response at the (x, y) position in the nth feature map, represents the value of the k1th feature response ranking of the (x, y) position in all feature maps obtained through the convolution layer, wherein the ranking is arranged in descending order of feature values; β r and β l are hyperparameters, representing the excitation coefficient and the inhibition coefficient respectively.

4. The image classification method based on hierarchical competitive pulse neural network according to claim 1 is characterized in that: The method of sorting different spiking neurons based on sensitivity and using the sorting result to stimulate or inhibit the spiking neurons includes: Using the second mapping function, the k2 spiking neurons that are most sensitive to the input information retain their original responses, the spiking neurons whose sensitivity to the input information is ranked after k2 and whose sensitivity is above 0 are inhibited from responding, and the spiking neurons that have no response to the input information are set to 0.

5. The image classification method based on hierarchical competitive pulse neural network according to claim 4 is characterized in that: The following formula is used to implement the sorting of different spiking neurons based on sensitivity, and the sorting result is used to implement the stimulation or inhibition of the spiking neurons: Among them, t represents the time step in the SNN network, and the second mapping function o t,m represents the response value output by the mth pulse neuron after mapping, f t,m represents the output response value of the mth spiking neuron, f t,topk It represents the k2th response value obtained by sorting the response values ​​of each pulse neuron in the fully connected layer from large to small.

6. The image classification method based on hierarchical competitive pulse neural network according to claim 1 is characterized in that: The step of assigning weights to the neurons in the fully connected layer includes: During the back-propagation process, the global learning weights, local learning weights, and weight distribution ratio between the global learning weights and the local learning weights of the neurons are calculated.

7. The image classification method based on hierarchical competitive pulse neural network according to claim 1 or 6, characterized in that: The weights of each neuron in the fully connected layer are calculated using the following formula: Among them, ∝ is a hyperparameter, which represents the weight distribution ratio, represents the global learning weight of the corresponding neuron, represents the local learning weight of the corresponding neuron, l represents the number of spiking neurons in the fully connected layer, and N represents the number of convolutional layers.

8. The image classification method based on hierarchical competitive pulse neural network according to claim 1 is characterized in that: The training of the SNN network also includes: The STDP rule is adopted to utilize the pulse correlation of the pulse neurons in the front and back layers to enhance or weaken the synaptic weight, including: when the pulse correlation of the pulse neurons in the front and back layers has a temporal causal relationship, the synaptic connection is enhanced; otherwise, the synaptic connection is weakened.

9. An image classification system based on a hierarchical competitive spiking neural network, characterized in that: It includes image acquisition module, image preprocessing module, SNN network module and training module, among which: Image acquisition module: used to acquire images; Image preprocessing module: used to preprocess the collected image data; SNN network module: used to calculate the predicted classification label of the image using the preprocessed image data; the SNN network module introduces a competition strategy, including: Passing the input image data through a convolution layer to obtain a feature map, sorting the feature response of each position in the feature map, and using the sorting result to achieve excitation or inhibition of the feature map, wherein the magnitude of the excitation or inhibition is obtained by the excitation coefficient and the inhibition coefficient respectively; and sorting different pulse neurons based on sensitivity, and using the sorting result to achieve excitation or inhibition of the pulse neurons; Training module: used to assign weights to each convolution kernel in the convolution layer of the SNN network and to each neuron in the fully connected layer; calculate the loss value of the SNN network training; use the loss value for back propagation to calculate the weights of each convolution kernel, the weights of each neuron, the excitation coefficient and the inhibition coefficient.

10. The image classification system based on hierarchical competitive spiking neural network according to claim 9, characterized in that: The weights of each neuron in the fully connected layer are calculated using the following formula: Among them, ∝ is a hyperparameter, which represents the weight distribution ratio, represents the global learning weight of the corresponding neuron, represents the local learning weight of the corresponding neuron, l represents the number of spiking neurons in the fully connected layer, and N represents the number of convolutional layers.

Images