Quantum-derived spiking convolutional neural network, image processing method and system
By introducing quantum-derived convolutional computing layer and quantum-heurized neurons into pulsed convolutional neural networks, the problems of negative value processing and bias update are solved, the efficiency and robustness of the network are improved, and the structure and resource use are optimized.
Patent Information
- Application Number
- CN202311636117.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-11-30
- Publication Date
- 2024-12-17
- Estimated Expiration
- 2043-11-30
AI Technical Summary
Existing pulse convolutional neural networks lose information when processing negative values, resulting in increased errors, and bias indicating negative values increases hardware resources and power consumption. The existing solution resets the bias after each training iteration Setting will also increase network error.
The convolution layer is replaced by quantum-derived convolutional operation layer, and the ReLU activation layer is removed, and the fully connected layer is replaced by quantum-heurized neurons. The data is processed through the principles of quantum superposition and quantum entanglement, so that the data is non-negative, enhancing the nonlinearity of the network. Processing capability and conditional update bias at each update iteration to reduce error.
It effectively improves the efficiency and robustness of pulse convolution deep neural networks, reduces network errors, optimizes network structure, and reduces hardware resources and power consumption.
Smart Images
Figure CN117574968B_ABST
Abstract
Description
Technical Field
[0001] The present disclosure relates to the technical field related to quantum models, and more specifically, to a quantum-derived pulse convolutional neural network, image processing method and system. Background Art
[0002] The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute prior art.
[0003] Spiking neural networks (SNNs): SNNs are models that use the biological neuron mechanism that is most suitable for calculation. They are closer to the working mechanism of the human brain than current deep learning techniques. Convolutional Neural Networks (CNNs) such as Figure 1 As shown in, it consists of one or more convolutional layers, ReLU activation layers, pooling layers and fully connected layers; the pulse convolutional neural network is as follows Figure 2 As shown in the figure, it is a new method to convert CNN into SNN, which can map CNN to a peak-based hardware architecture, that is, setting an absolute value (Abs) function layer and a pulse generation layer at the front end of the convolutional neural network, and connecting a spike counter to the fully connected layer of the convolutional neural network; transforming the ordinary CNN architecture into a customized CNN architecture to adapt to the requirements of SNN, and then training the customized CNN in the same way as CNN, and finally applying the learned network weights to the SNN architecture derived from the customized CNN.
[0004] Quantum derivative technology is a high-performance information processing technology with quantum informatics characteristics implemented on existing electronic computer hardware systems by improving classical information processing technology based on quantum information principles such as quantum superposition and quantum entanglement. Its high performance is mainly reflected in its good nonlinear processing capabilities. Quantum derivative technology is mainly developed in combination with specific fields such as artificial intelligence, signal processing, and image processing. Among them, the introduction of quantum derivative technology into artificial intelligence technology has led to the emergence of research on quantum derivative neural networks, quantum derivative genetic algorithms, and quantum derivative swarm intelligence technologies.
[0005] The inventors found in their research that the existing spiking neural network model has the following problems:
[0006] First, since it is difficult to represent negative values in spikes, an absolute value layer is added to the spike convolutional neural network, namely Figure 2 The Abs function layer in the , this operation will lose some negative input information, which will increase the error of the network;
[0007] Second, in the spike neuron stage, since the bias representing negative values will bring more hardware resources and higher power consumption, the existing practice is to reset all bias values to zero after each training iteration, which will also increase the network error. Summary of the invention
[0008] In order to solve the above problems, the present invention proposes a quantum-derived pulse convolutional neural network, an image processing method and a system, which can be used to realize applications such as image classification and target detection, and can effectively improve the efficiency of the pulse convolutional deep neural network and enhance the robustness of the pulse convolutional deep neural network.
[0009] In order to achieve the above objectives, the present disclosure adopts the following technical solutions:
[0010] One or more embodiments provide a quantum-derived spiking convolutional neural network, comprising a preprocessing module, a spiking generation layer, a spiking neuron, and a spiking counter connected in sequence;
[0011] The spiking neuron includes one or more convolution groups, quantum-derived convolution operation layers, and quantum-inspired neurons; a convolution group includes a convolution layer, a ReLU activation layer, and a pooling layer;
[0012] A quantum-derived convolution operation layer is used to replace the convolution layer in one or more convolution groups of the pulse convolutional neural network, and the ReLU activation layer connected after the quantum-derived convolution operation layer is removed. The quantum-derived convolution operation layer performs the quantum-derived convolution operation.
[0013] One or more embodiments provide an image processing method based on the above-mentioned quantum-derived spiking convolutional neural network, comprising the following steps:
[0014] Obtain image data to be classified and preprocess the input image data;
[0015] Detecting image pixels of the preprocessed image, and generating a pulse sequence when a set condition is met;
[0016] Perform quantum-derived convolution, convolution, activation and spatial linear subsampling operations on the input pulse sequence data, transmit it to the quantum-inspired neuron for conversion, and obtain output pulses;
[0017] The pulse peaks of all categories of neurons are counted, and the category output corresponding to the neuron with the largest number of peaks is obtained as the classification result of the image.
[0018] One or more embodiments provide an image processing system based on a quantum-derived spike convolutional neural network, comprising: an image acquisition device and a processor;
[0019] An image acquisition device is configured to acquire images to be classified, and specifically may be a camera device;
[0020] A processor is configured to execute the above-mentioned image processing method based on quantum-derived pulse convolutional neural network.
[0021] An electronic device includes a memory and a processor and computer instructions stored in the memory and running on the processor. When the computer instructions are run by the processor, the steps in the above-mentioned image processing method based on quantum-derived pulse convolutional neural network are completed.
[0022] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the steps in the above-mentioned image processing method based on quantum-derived pulse convolutional neural network.
[0023] Compared with the prior art, the present invention has the following beneficial effects:
[0024] The present invention introduces quantum-derived technology into pulse convolutional neural networks, and uses quantum-derived convolution operations and quantum-inspired neurons to replace convolution operations and fully connected layers, respectively. In the quantum-derived convolution operation, the input data is first converted into a quantum-like state, which not only ensures that the data is non-negative, but also enhances the nonlinearity of the data. On the other hand, by introducing quantum-derived technology, the nonlinearity of the network is enhanced, the ReLU function is removed, and the network structure is further optimized.
[0025] The advantages of the present disclosure and the advantages of additional aspects will be described in detail in the following specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] The accompanying drawings constituting a part of the present disclosure are used to provide a further understanding of the present disclosure. The exemplary embodiments of the present disclosure and the description thereof are used to explain the present disclosure but do not constitute a limitation of the present disclosure.
[0027] Figure 1 is a convolutional neural network architecture diagram described in the background technology of this disclosure;
[0028] Figure 2 is a pulse convolutional neural network architecture diagram described in the background technology of this disclosure;
[0029] Figure 3 This is a diagram of the quantum-derived pulse convolutional neural network architecture of Example 1 of the present disclosure;
[0030] Figure 4 is a structural diagram of a quantum-inspired neuron according to Example 1 of the present disclosure;
[0031] Figure 5 It is a schematic diagram of the quantum-derived convolution operation of Example 1 of the present disclosure. DETAILED DESCRIPTION
[0032] The present disclosure is further described below in conjunction with the accompanying drawings and embodiments.
[0033] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present disclosure belongs.
[0034] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present disclosure. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof. It should be noted that, in the absence of conflict, the various embodiments in the present disclosure and the features in the embodiments can be combined with each other. The embodiments will be described in detail below in conjunction with the accompanying drawings.
[0035] Example 1
[0036] In the technical solutions disclosed in one or more embodiments, Figures 2 to 5 As shown, the quantum-derived pulse convolutional neural network includes a preprocessing module, a pulse generation layer, a pulse neuron, and a pulse counter connected in sequence;
[0037] The spiking neuron includes one or more convolution groups, quantum-derived convolution operation layers, and quantum-inspired neurons; a convolution group includes a convolution layer, a ReLU activation layer, and a pooling layer;
[0038] A quantum-derived convolution operation layer is used to replace the convolution layer in one or more convolution groups of the pulse convolutional neural network, and the ReLU activation layer connected after the quantum-derived convolution operation layer is removed. The quantum-derived convolution operation layer performs the quantum-derived convolution operation.
[0039] In this embodiment, quantum derivation is combined with a pulse convolutional neural network to optimize the structure of the existing pulse convolutional neural network, greatly improving the performance of the network.
[0040] The CNN-based pulse convolutional neural network not only has the potential for low energy consumption, but also maintains high performance in image classification, etc. However, since a large amount of data needs to be discarded in the process of converting CNN into a pulse convolutional neural network, such as negative values in the input data and all biases in each layer of CNN, the data deviation is large.
[0041] In this embodiment, quantum-derived technology is introduced into the pulse convolutional neural network, and quantum-derived convolution operations and quantum-inspired neurons are used to replace convolution operations and fully connected layers, respectively. In the quantum-derived convolution operation, the input data is first converted into a quantum-like state, which not only ensures that the data is non-negative, but also enhances the nonlinearity of the data. On the other hand, by introducing quantum-derived technology, the network nonlinearity is enhanced, the ReLU function is removed, and the network structure is further optimized.
[0042] Furthermore, during the quantum-derived convolution operation, the value range of the convolution kernel weight and bias of the quantum-derived convolution operation layer is corrected so that the weight and bias values are non-negative.
[0043] Furthermore, during the data processing of the quantum-inspired neuron, the value range of the weight and bias of the quantum-inspired neuron is modified so that the value of the weight and bias becomes non-negative;
[0044] A preprocessing module is configured to preprocess the input data, including normalization, applying a phase, and converting into a quantum state;
[0045] In this embodiment, each time the iterative bias is updated, the bias is updated conditionally to keep a non-negative value, thereby reducing the network error.
[0046] A pulse generation layer is configured to detect image pixels of the preprocessed image and generate a pulse train when a set condition is met and transmit it to the pulse neuron;
[0047] The pulse neuron performs quantum-derived convolution, convolution, activation and spatial linear subsampling operations on the input pulse sequence data, and transmits them to the quantum-inspired neuron for conversion to obtain output pulses;
[0048] A pulse counter is configured to count the pulse peaks of all categories of neurons and obtain the category output corresponding to the neuron with the largest number of peaks;
[0049] Different pulse convolutional neural networks contain different numbers of convolutional layers and fully connected layers. The pulse convolutional neural network of this embodiment can adopt any number of convolution groups, use quantum-derived convolution operations and quantum-inspired neurons to replace different convolutional layers and fully connected layers respectively, and generate different quantum-derived pulse convolutional neural networks.
[0050] In this embodiment, specifically, Figure 2 The spiking convolutional neural network shown in Figure 1 is composed of two convolutional groups (convolutional layer, ReLU activation layer and pooling layer) and one fully connected layer.
[0051] The various parts of the quantum-derived spiking convolutional neural network are explained below.
[0052] Data processing module,In quantum computing, a qubit is implemented using a two-level quantum system.,It can be described by a two-dimensional complex Hilbert space.
[0053] According to the basic principle of quantum state superposition, any quantum state can be written as α and β are the probability amplitudes of the calculation bases |0> and |1>, respectively. The amplitudes α and β are complex numbers that satisfy the normalization condition, i.e., |α| 2 +|β| 2 = 1. Therefore, according to the superposition state and normalization conditions, a quantum heuristic of input data can be expressed as:
[0054] f(θ)=e iθ =cosθ+i sinθ
[0055] Input data preprocessing includes the following steps:
[0056] 1) Normalization: Perform spatial normalization on the input image data to obtain a value between [0,1];
[0057] 2) Apply phase: Multiply the normalized data by π / 2 to obtain the phase of the complex input in the range [0,π / 2].
[0058] 3) Convert to quantum state: Quantum-inspired encoding of the input data after applying the phase can be expressed as:
[0059]
[0060] x real =cos(x input1 ),x imag = sin(x input1 )
[0061] Among them, x data is the normalized input data; x input1 is the data after applying the phase; x real is the real part of the quantum state data after transformation, x imag is the imaginary part of the quantum state data after transformation; x real and x imag are all positive real numbers between [0,1].
[0062] The pulse generation process of the pulse generation layer is consistent with the pulse convolutional neural network;
[0063] In the pulse neuron of this embodiment, the convolution and full connection operations in the pulse neuron are replaced by quantum-derived convolution operations and quantum-inspired neurons, respectively.
[0064] In the spiking neuron, the quantum-derived convolution operation of the quantum-derived convolution operation layer is as follows: Figure 5 As shown, the following steps are included:
[0065] 1) Set the convolution kernel weight w 10 and 20 , bias b 10 and b 20 The initial value is assigned to the standard normal distribution, and the full connection weight w f and bias b f Assign the initial value to uniform distribution;
[0066] 2) Calculate the real and imaginary parts of the quantum-derived convolution kernel weights and the real and imaginary parts of the bias corresponding to the convolution operation respectively;
[0067] The real part of the quantum derived convolution kernel weight w real1 and the imaginary part w imag1 They are:
[0068] w real1 =cos(w 10 ),w imag1 = sin(w 10 );
[0069] The real part b of the bias corresponding to the convolution operation real1 and the imaginary part b imag1 They are:
[0070] b real1 =cos(b 10 ),b imag1 = sin(b 10 );
[0071] Among them, w real1 、w imag1 , b real1 , b imag1 All are non-negative values;
[0072] 3) According to the obtained convolution kernel weight and bias, the quantum derivative convolution operation is performed as follows:
[0073]
[0074] Then perform a controlled NOT operation to obtain the feature y extracted by the convolution operation. c_1 ,Right now:
[0075]
[0076] y c_1 The real and imaginary data are stored in different channels respectively, and the channels are merged.
[0077] Based on the good nonlinearity of quantum-derived convolution operation, the ReLU function is omitted after the quantum-derived convolution operation layer in this embodiment; spatial linear subsampling is directly performed based on the maximum pooling layer to obtain y pool1 .
[0078] In the spiking neurons, the classical fully connected layers are replaced with quantum-inspired neurons, which are configured to perform the following process:
[0079] 1) The output of the connected convolutional group is normalized and converted into a quantum state through quantum-inspired encoding, namely:
[0080]
[0081] Among them, y pool2 is the normalized data; input is the obtained quantum state data;
[0082] 2) Use quantum-inspired neurons to replace the fully connected layer, perform quantum gate operations, and obtain the output of quantum-inspired neurons;
[0083] υ1=f(w f )y input -f(b f ), y output1 =f(θ1),
[0084] Among them, w f is the weight of the fully connected layer, b f is the bias of the fully connected layer.
[0085] The final output is y output =|Im(y output1 )| 2 .
[0086] Specifically, the data processing process of quantum-inspired neurons is as follows: Figure 4 As shown, we first need to quantize the input data; then apply quantum gate operations such as rotation gate and controlled NOT gate to the input data. N}, normalize the input data {x 01 ,...x 0N}, firstly perform quantum inspired coding, that is, Then implement the revolving door, that is, the weighted sum Finally, construct the controlled NOT gate parameters Executing the controlled NOT gate gives the output S=f(θ) of the quantum-inspired neuron.
[0087] The pulse counter, based on the pulse counting of the quantum-derived pulse convolutional neural network, is consistent with the pulse convolutional neural network. Starting from the input image, within a set time period (such as 100 or 200 time steps), the pulse peaks of all categories of neurons in the linear classification layer are counted; then the category output corresponding to the neuron with the largest number of peaks is generated, which is the classification result.
[0088] Example 2
[0089] Based on Example 1, this embodiment provides an image processing method based on the quantum-derived pulse convolutional neural network described in Example 1, inputting the image to be processed into the quantum-derived pulse convolutional neural network described in Example 1 to achieve image classification, including the following steps:
[0090] Step 1: Obtain image data to be classified and preprocess the input image data;
[0091] Step 2, detecting image pixels of the preprocessed image, and generating a pulse sequence when a set condition is met;
[0092] Step 3: Perform quantum-derived convolution operation, convolution operation, activation operation and spatial linear subsampling operation on the input pulse sequence data, and transmit it to the quantum-inspired neuron for conversion to obtain output pulses;
[0093] Step 4: Count the pulse peaks of all categories of neurons, and obtain the category output corresponding to the neuron with the largest number of peaks as the classification result of the image;
[0094] In step 1, data preprocessing is implemented in the preprocessing module, including normalization, applying phase, and converting to quantum state, including the following processes:
[0095] 1) Normalization: Perform spatial normalization on the input image data to obtain a value between [0,1];
[0096] 2) Apply phase: Multiply the normalized data by π / 2 to obtain the phase of the complex input in the range [0,π / 2].
[0097] 3) Convert to quantum state: Quantum-inspired encoding of the input data after applying the phase can be expressed as:
[0098]
[0099] x real =cos(x input1 ),x imag = sin(x input1 )
[0100] Among them, x datais the normalized input data; x input1 is the data after applying the phase; x real is the real part of the quantum state data after transformation, x imag is the imaginary part of the quantum state data after transformation;
[0101] In step 2, the pulse generation layer is implemented. The pulse generation process is consistent with the pulse convolutional neural network and is defined as: Assume x ijk (k=1,2,3) is the image map, which is used as the input of the pulse generation layer. At time t, when the neuron (i,j) in the kth channel satisfies rand() <c x ijk When , a pulse is generated, where rand() is a random number generator uniformly distributed on (0,1), and c is a constant used to scale the frequency of generated pulses.
[0102] In step 3, the input pulse sequence data is subjected to quantum-derived convolution operation, convolution operation, activation operation and spatial linear subsampling operation, and output to the quantum-inspired neuron for conversion to obtain output pulses;
[0103] Among them, the execution order and number of executions of the quantum-derived convolution operation, convolution operation, activation operation and spatial linear subsampling operation are determined according to the structure of the quantum-derived pulse convolutional neural network described in Example 1 and the number of convolution groups.
[0104] Take two groups of convolution as an example, Figure 3 As shown, quantum derivative convolution operation, maximum pooling, ordinary convolution, ReLU activation operation, maximum pooling, and spatial linear subsampling are performed on the input pulse sequence data in sequence.
[0105] The quantum-derived convolution operation is implemented in the quantum-derived convolution operation layer and includes the following steps:
[0106] 1) Convolution kernel weight w 10 and 20 , bias b 10 and b 20 The initial value is assigned to the standard normal distribution, and the full connection weight w f and bias b f Assign the initial value to uniform distribution;
[0107] 2) Calculate the real and imaginary parts of the quantum-derived convolution kernel weights and the real and imaginary parts of the bias corresponding to the convolution operation respectively;
[0108] The real part of the quantum derived convolution kernel weight w real1 and the imaginary part w imag1 They are:
[0109] w real1=cos(w 10 ),w imag1 = sin(w 10 );
[0110] The real part b of the bias corresponding to the convolution operation real1 and the imaginary part b imag1 They are:
[0111] b real1 =cos(b 10 ),b imag1 = sin(b 10 );
[0112] Among them, w real1 、w imag1 , b real1 , b imag1 All are non-negative values;
[0113] 3) According to the obtained convolution kernel weight and bias, the quantum derivative convolution operation is performed as follows:
[0114]
[0115] Then perform a controlled NOT operation to obtain the feature y extracted by the convolution operation. c_1 ,Right now:
[0116]
[0117] y c_1 The real and imaginary data are stored in different channels respectively, and the channels are merged.
[0118] Based on the good nonlinearity of quantum-derived convolution operation, the ReLU function is omitted after the quantum-derived convolution operation layer in this embodiment; spatial linear subsampling is directly performed based on the maximum pooling layer to obtain y pool1 .
[0119] Quantum-inspired neurons are configured to perform the following processes:
[0120] 1) The output of the connected convolutional group is normalized and converted into a quantum state through quantum-inspired encoding, namely:
[0121]
[0122] Among them, y pool2 is the normalized data; input is the obtained quantum state data;
[0123] 2) Use quantum-inspired neurons to replace the fully connected layer, perform quantum gate operations, and obtain the output of quantum-inspired neurons;
[0124] υ1=f(w f )y input -f(b f ), y output1 =f(θ1),
[0125] Among them, w f is the weight of the fully connected layer, b f is the bias of the fully connected layer.
[0126] The final output is y output =|Im(y output1 )| 2 .
[0127] The pulse count of the quantum-derived spiking convolutional neural network is consistent with that of the spiking convolutional neural network. Starting from the input image, the spike peaks of all categories of neurons in the linear classification layer are counted within a set time period (such as 100 or 200 time steps); then the category output corresponding to the neuron with the largest number of spikes is generated, which is the classification result.
[0128] Example 3
[0129] Based on Example 1, this embodiment provides an image processing system based on a quantum-derived pulse convolutional neural network, including: an image acquisition device and a processor;
[0130] An image acquisition device is configured to acquire images to be classified, and specifically may be a camera device;
[0131] The processor is configured to execute the image processing method described in Example 2.
[0132] Example 4
[0133] This embodiment provides an electronic device, including a memory and a processor, and computer instructions stored in the memory and running on the processor. When the computer instructions are executed by the processor, the steps in the image processing method based on quantum-derived pulse convolutional neural network of Example 2 are completed.
[0134] Example 5
[0135] This embodiment provides a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, the steps in the image processing method based on quantum-derived pulse convolutional neural network of Example 2 are completed.
[0136] The above description is only a preferred embodiment of the present disclosure and is not intended to limit the present disclosure. For those skilled in the art, the present disclosure may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present disclosure shall be included in the protection scope of the present disclosure.
[0137] Although the above describes the specific implementation methods of the present disclosure in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present disclosure. Technical personnel in the relevant field should understand that on the basis of the technical solution of the present disclosure, various modifications or variations that can be made by those skilled in the art without creative work are still within the scope of protection of the present disclosure.
Claims
1. Image processing method based on quantum-derived spiking convolutional neural network, It is characterized in that The steps include: Obtain image data to be classified and preprocess the input image data; Detecting image pixels of the preprocessed image, and generating a pulse sequence when a set condition is met; Perform quantum-derived convolution, convolution, activation and spatial linear subsampling operations on the input pulse sequence data, transmit it to the quantum-inspired neuron for conversion, and obtain output pulses; Count the pulse peaks of all categories of neurons, and obtain the category output corresponding to the neuron with the largest number of peaks as the classification result of the image; The steps of the image processing method are implemented by a quantum-derived pulse convolutional neural network, which includes a preprocessing module, a pulse generation layer, a pulse neuron, and a pulse counter connected in sequence; The input of the spike generation layer is the image map x ijk (k=1,2,3), at time t, when the neuron (i,j) in the kth channel satisfies rand() <c x ijk When , a pulse is generated, where rand() is a random number generator uniformly distributed on (0,1), and c is a constant used to scale the frequency of the generated pulse; The spiking neuron includes one or more convolution groups, quantum-derived convolution operation layers, and quantum-inspired neurons; a convolution group includes a convolution layer, a ReLU activation layer, and a pooling layer; Replacing one or more convolutional layers of the pulse convolutional neural network with a quantum-derived convolutional operation layer, and removing the ReLU activation layer connected after the quantum-derived convolutional operation layer, wherein the quantum-derived convolutional operation layer performs the quantum-derived convolution operation; The quantum-derived convolution operation of the quantum-derived convolution operation layer includes the following steps: Set the convolution kernel weight w 10 and 20 , bias b 10 and b 20 Assign the initial value to the standard normal distribution and set the full connection weight w f and bias b f Assign the initial value to uniform distribution; Calculate the real and imaginary parts of the quantum derived convolution kernel weights and the real and imaginary parts of the bias corresponding to the convolution operation respectively; The real part of the quantum derived convolution kernel weight w real1 and the imaginary part w imag1 They are: w real1 =cos(w 10 ),w imag1 =sin(w 10 ); The real part b of the bias corresponding to the convolution operation real1 and the imaginary part b imag1 They are: b real1 =cos(b 10 ),b imag1 =sin(b 10 ); Among them, w real1 、w imag1 、b real1 、b imag1 All are non-negative values; According to the obtained convolution kernel weight and bias, a quantum-derived convolution operation is performed; In the spiking neurons, the fully connected layers are replaced with quantum-inspired neurons, which are configured to perform the following process: The output of the connected convolutional group is normalized and converted into a quantum state through quantum-inspired encoding, namely: in, ypool2 is the normalized data; yinput is the obtained quantum state data; Perform quantum gate operations to obtain the output of quantum-inspired neurons; Among them, w f is the weight of the fully connected layer, b f is the bias of the fully connected layer; The final output is y output =|Im(y output1 )| 2 ; During the quantum-derived convolution operation, the value range of the convolution kernel weight and bias of the quantum-derived convolution operation layer is modified so that the weight and bias values are non-negative; During the data processing by quantum-inspired neurons, the value ranges of weights and biases of quantum-inspired neurons are corrected so that the values of weights and biases become non-negative.
2. Image processing system based on quantum-derived spiking convolutional neural network, It is characterized in that include: Image acquisition device and processor; An image acquisition device, configured to acquire images to be classified, specifically a camera device; A processor configured to execute the image processing method based on quantum-derived pulse convolutional neural network as described in claim 1.
3. An electronic device, It is characterized in that The invention comprises a memory and a processor and computer instructions stored in the memory and executed on the processor. When the computer instructions are executed by the processor, the steps of the image processing method based on quantum-derived pulse convolutional neural network described in claim 1 are completed.
4. A computer-readable storage medium, It is characterized in that Used to store computer instructions, which, when executed by a processor, complete the steps of the image processing method based on quantum-derived pulse convolutional neural network described in claim 1.