Quantum generative adversarial network optimization method based on radial basis network

Abstract

The invention discloses an image processing method and system based on RBF-QGAN, and belongs to the field of quantum generation models. The method comprises the following steps: acquiring image data, and preprocessing the image data to obtain discrete sample data; an RBF-QGAN comprising a discriminator and a generator is constructed, the discriminator adopts a radial basis network, and the generator adopts a variable component sub-circuit; the generator receives discrete samples, learns an input generation probability distribution, and generates an output probability distribution using a quantum circuit. The discriminator receives the data output by the generator and judges the authenticity; generating probability distribution and discrete sample data which are learned and generated by the generator are input into a discriminator, feature extraction is carried out on the samples through a radial basis network, and the authenticity of the data samples is judged; training is carried out through a cross entropy loss function, parameter optimization is carried out through an Adam optimizer, and then the parameters are transmitted into a generator and a discriminator for training until training loss converges to Nash equilibrium; and after optimization training is finished, outputting an image data sample generated by the generator.

Description

Technical Field

[0001] The present invention belongs to the field of quantum generative models, and specifically relates to an image processing method and system based on RBF-QGAN. Background Art

[0002] Generative adversarial networks (GANs) are an unsupervised learning algorithm designed to solve generative modeling problems. They have attracted widespread attention for their outstanding performance in generating high-quality data. Generative adversarial networks provide a way to learn deep representations without a large amount of annotated training data. GANs were proposed by Goodfellow et al. in 2014 and are generative models constructed in a completely new way. The core idea is to train two neural networks, the generator and the discriminator, in an adversarial way, so that the generator can generate realistic data, and the discriminator is used to judge the difference between the generated data and the real data. In the process of mutual competition, the generator is constantly improved, thereby generating data of higher and higher quality. Since its proposal, GANs have been widely used in image generation, data enhancement, image editing, text generation and other fields, and have developed multiple variants such as conditional GANs, CycleGANs, and StyleGANs.

[0003] However, despite the great success of traditional machine learning in many aspects, it still faces many challenges. As the complexity of the model increases, traditional machine learning algorithms and deep learning models face the limitation of computing power bottlenecks. Especially when dealing with massive data and ultra-high-dimensional tasks, the computational cost of training and inference rises sharply, which limits the development of more complex models and the popularization of large-scale applications.

[0004] Faced with this bottleneck, the rise of quantum computing provides a possible solution. Quantum computing uses the basic principles of quantum mechanics, such as superposition, entanglement, and quantum interference, and has parallel processing capabilities that surpass classical computers and exponential computing acceleration potential, which makes it show unique advantages in solving computationally intensive problems. Therefore, combining quantum computing with machine learning and developing quantum machine learning (QML) is considered to be one of the innovative directions to solve the computing power bottleneck. With the continuous development of quantum hardware technology, the practical application prospects of quantum computers are becoming increasingly broad. Against this background, quantum machine learning, as a research direction in the intersection of quantum computing and machine learning, has gradually become the focus of academic and industrial attention. Quantum machine learning, that is, the integration of machine learning and quantum computing, is gradually becoming a key application of quantum technology. At present, researchers have long begun to explore the combination of machine learning algorithms with quantum capabilities and have achieved impressive results.

[0005] As a deep learning model, generative adversarial networks require more computing power from classical computers than ordinary neural networks, so the need to introduce quantum computing into traditional GANs is more urgent. Quantum generative adversarial networks combine the advantages of quantum computing and GANs. By utilizing the parallel computing power of quantum computing and the high-dimensional representation ability of quantum states, they are expected to make breakthrough progress in solving the problems faced by classical GANs. QGANs can not only improve the quality and diversity of generated data, but also show the potential to surpass classical methods in certain specific tasks. However, the quantum generative adversarial networks used today still have problems such as unstable training and poor robustness to training noise, and new solutions are urgently needed. Summary of the invention

[0006] In view of the deficiencies in the prior art, the object of the present invention is to provide an image processing method and system based on RBF-QGAN, which solves the problems in the prior art.

[0007] The purpose of the present invention can be achieved through the following technical solutions:

[0008] An image processing method based on RBF-QGAN, comprising the following steps:

[0009] Get image data;

[0010] Preprocessing the acquired image data to obtain discrete sample data;

[0011] Construct an RBF-QGAN consisting of a discriminator and a generator. The discriminator uses a radial basis network and the generator uses a variational quantum circuit. The generator receives discrete samples, learns the input generation probability distribution, and uses quantum circuits to generate output probability distributions. The discriminator receives the data output by the generator and determines the authenticity of the data.

[0012] The generated probability distribution learned by the generator and the preprocessed discrete sample data are input into the discriminator, and the features of the samples are extracted through the radial basis network to determine the authenticity of the data samples;

[0013] The cross entropy loss function is used for training and the Adam optimizer is used for parameter optimization. The optimized parameters are then passed into the generator and discriminator for further training until the training losses of the generator and discriminator converge to the Nash equilibrium.

[0014] After the optimization training is completed, the image data samples generated by the generator are output.

[0015] Furthermore, the step of image data preprocessing includes:

[0016] Reduce the image data to image data with a dimension of 16×16;

[0017] The image data after dimensionality reduction is blurred and normalized;

[0018] The processed image data is converted into discrete sample data as the true distribution.

[0019] Furthermore, the radial basis network includes: an input layer, a hidden layer and an output layer; the radial basis network uses a radial basis function as a Gaussian kernel function, which is expressed as:

[0020] H(x,c)=e(-β·‖xc‖ 2 )

[0021] Among them, ||·|| is the norm on the function space, x i represents the i-th sample in the training set, x is the input data, c is the center of the RBF layer, and β is defined as the expansion coefficient.

[0022] Furthermore, the output y of the radial basis network i for:

[0023]

[0024] Where ω0 represents the bias of the output neuron, q represents the number of hidden neurons, and w j represents the weight between the jth hidden neuron and the output neuron; v j and σ j They represent the center and width of the jth hidden neuron respectively; g1 represents the Gaussian kernel function.

[0025] Furthermore, the variational quantum circuit includes a rotation layer and an entanglement layer; the rotation layer uses a single quantum bit Ry gate; the entanglement layer uses a CZ gate, which is connected in a ring topology.

[0026] An image processing system based on RBF-QGAN, comprising:

[0027] Data acquisition module: acquire image data;

[0028] Data preprocessing module: preprocess the acquired image data to obtain discrete sample data;

[0029] Model building module: Build an RBF-QGAN including a discriminator and a generator. The discriminator uses a radial basis network and the generator uses a variational quantum circuit. The generator receives discrete samples, learns the input generation probability distribution, and uses quantum circuits to generate output probability distributions. The discriminator receives the data output by the generator and determines the authenticity of the data.

[0030] Adversarial training module: The generated probability distribution learned by the generator and the preprocessed discrete sample data are input into the discriminator, and the features of the samples are extracted through the radial basis network to determine the authenticity of the data samples;

[0031] Parameter optimization module: trains through the cross entropy loss function and optimizes parameters through the Adam optimizer, then passes the optimized parameters to the generator and discriminator for further training until the training losses of the generator and discriminator converge to the Nash equilibrium;

[0032] And, data sample generation module: after the optimization training is completed, the image data samples generated by the output generator are output.

[0033] A computer storage medium stores a readable program, and when the program is run, the above-mentioned image processing method based on RBF-QGAN can be executed.

[0034] An electronic device, comprising: a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other through the communication bus;

[0035] The memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform an operation corresponding to the above-mentioned image processing method based on RBF-QGAN.

[0036] A computer program product includes computer instructions, wherein the computer instructions instruct a computing device to perform operations corresponding to the above-mentioned image processing method based on RBF-QGAN.

[0037] A hybrid quantum generative adversarial network based on a radial basis network includes a discriminator and a generator. The discriminator adopts a radial basis network. The generator receives discrete samples of input image data, learns the input generation probability distribution, and uses a quantum circuit to generate an output probability distribution; the discriminator receives the data output by the generator and judges the authenticity of the data.

[0038] Beneficial effects of the present invention:

[0039] 1. The present invention uses a quantum classical generative adversarial network with a radial basis network as a discriminator and a variational quantum circuit as a generator; by using this hybrid structure, it allows the advantages of quantum computing to be combined with classical deep learning, which helps to make full use of the capabilities of classical computers and may be more effective in solving certain problems.

[0040] 2. This invention not only effectively solves problems such as scale, training stability and generation quality, but also brings new possibilities for the application of quantum generative adversarial networks in the field of artificial intelligence, which is currently facing a computing power bottleneck.

[0041] 3. This paper overcomes the limitations of the existing technology, such as unstable training and easy collapse, insufficient training noise robustness, and poor generation quality, thereby significantly enhancing the ability of HQCGAN to generate high-fidelity grayscale images with discrete value distribution. Through careful experiments, evaluation of training cross-validation scores and loss function robustness, the excellent performance of the RBF-QGAN model is demonstrated, especially in the presence of noisy input data. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0043] Figure 1 It is a schematic diagram of the overall structure of the hybrid quantum classical generative adversarial network of the present invention;

[0044] Figure 2 It is a schematic diagram of the structure of a hybrid quantum generative adversarial network based on a radial basis network of the present invention;

[0045] Figure 3 It is a schematic diagram of the radial basis network structure of the present invention;

[0046] Figure 4 is a schematic diagram of a quantum generation circuit of the present invention;

[0047] Figure 5 It is a schematic diagram of CV values ​​of various network models on the MNIST data set with different types of noise in the present invention;

[0048] Figure 6 The model of the present invention performs MNIST image data training and generates corresponding loss value change graphs;

[0049] Figure 7 This is a diagram of the robustness exploration experiment results of the present invention. DETAILED DESCRIPTION

[0050] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] Example 1

[0052] An image processing method based on RBF-QGAN, comprising the following steps:

[0053] S1, obtain image data; construct an image data set, or use an existing data set for image generation, such as MNIST, Fashion MNIST, CIFAR-10, etc., which is not specifically limited in the present invention. In this embodiment, the training data set uses the MNIST grayscale image data set.

[0054] S2, preprocessing the acquired image data to obtain discrete sample data;

[0055] The steps of image data preprocessing include:

[0056] S21, in order to adapt the quantum circuit to use an appropriate number of quantum bits, the image data is reduced in dimension to image data with a dimension of 16×16.

[0057] The dimensionality reduction method uses the interpolation scaling method to generate a scaled image by weighted averaging the pixel values ​​of the image, thus changing the size of the image. This method is relatively simple to calculate and easy to implement. Compared with the traditional dimensionality reduction PCA method, it is more suitable for dimensionality reduction of image data sets with larger data sets.

[0058] S22, performing fuzzy processing and normalization processing on the image data after dimensionality reduction;

[0059] Among them, blurring of image data can improve data quality, reduce noise, and ensure that the data is within a range suitable for model training. Use Gaussian blur to smooth image data and reduce the discreteness of the image, making it more suitable as a probability distribution. The blurred data is normalized to a probability distribution, with the goal of adjusting the sum of all pixel values ​​to 1 to form a probability distribution. Normalization is necessary because quantum circuits are initialized based on probability amplitudes.

[0060] The normalization formula is:

[0061]

[0062] Among them I ij is the image pixel value, i, j represents the i-th row and j-th column in the image.

[0063] S23, converting the processed image data into discrete sample data as the true distribution;

[0064] Specifically, sampling is performed from the normalized probability distribution, converted to binary representation, and the result is mapped to -1 and 1 to generate discrete samples suitable for the quantum generator as the input of RBF-QGAN.

[0065] S3, constructed radial basis network-based hybrid quantum generative adversarial network (RBF-QGAN);

[0066] Among them, the hybrid quantum classical generative adversarial network (HQCGAN) is Figure 1 As shown, when processing classical data input in HQCGAN, the input is encoded into a quantum state, which is essentially a nonlinear mapping from a low-dimensional Hilbert space to a high-dimensional space. The present invention selects a simple angle encoding with only a single layer of depth to alleviate the input bottleneck. Based on this, the present invention constructs a hybrid quantum generative adversarial network based on radial basis network (RBF-QGAN);

[0067] like Figure 2 As shown in the figure, RBF-QGAN includes a discriminator and a generator. The discriminator uses a radial basis network (RBFNN) and the generator uses a variational quantum circuit. The generator uses a variational quantum circuit. The generator receives discrete samples of input image data, learns the input generation probability distribution, and uses quantum circuits to generate output probability distributions. The discriminator receives the classical data output by the generator and determines whether these data come from the real data set. The optimization goal of the discriminator is to maximize its accuracy to distinguish between real data and generated data. Through adversarial training with the generator, the discriminator continuously improves its classification ability, while also prompting the generator to generate higher quality samples. During the entire adversarial training process, both the generator and the discriminator use the Adam optimizer.

[0068] like Figure 3 As shown, the radial basis network includes:

[0069] The radial basis network as a regression model is a three-layer feedforward neural network consisting of an input layer, a hidden layer, and an output layer. It should be noted that in the specific context of combining RBFNN with QGANs in the present invention, the number of output neurons is set to 1.

[0070] Compared with the fully connected network (FCN) using sigmoid or ReLU activation function, RBFNN uses radial basis function (RBF). It is worth noting that using Gaussian kernel function to map data to infinite dimensional space can be considered as a feature extraction method for input data. Gaussian kernel function is the most widely used radial basis function, which is expressed as:

[0071] H(x,c)=e(-β·‖xc‖ 2 )

[0072] where ||·|| is the norm on the function space, x irepresents the i-th sample in the training set, x is the input data, c is the center of the RBF layer, and β is defined as the expansion coefficient. By multiplying the distance by the scalar coefficient β, we can control how fast the function decays. Therefore, a higher β means a greater decline.

[0073] The kernel function of the radial basis network uses the Gaussian kernel function, and the output y of the radial basis network is i It can be written as:

[0074]

[0075] Where ω0 represents the bias of the output neuron, q represents the number of hidden neurons, and w j represents the weight between the jth hidden neuron and the output neuron; x i represents the i-th sample in the training set, v j and σ j They represent the center and width of the jth hidden neuron respectively; g1 represents the Gaussian kernel function.

[0076] In theory, RBFNN has multiple advantages over FC neural networks. First, the nonlinear mapping ability of RBFNN stems from its use of radial basis functions, which exhibit strong nonlinear mapping capabilities. Especially in fields characterized by complex patterns and nonlinear relationships, RBFNN may be superior to fully connected networks in capturing and modeling complex data relationships. Second, local representation and learning are inherent to RBFNN because each radial basis function responds more strongly to a specific local region of the input data. This inherent property makes RBFNN potentially better at processing local features and non-uniformly distributed data without considering the entire dataset. Third, due to the sensitivity of radial basis functions to local input data regions, robustness and generalization are significant advantages of RBFNN. This sensitivity helps to enhance robustness to noise and outliers, and ultimately improves the generalization ability of the model, especially in processing unseen data. Finally, RBFNN has the characteristics of fewer parameters and faster training speed, because it generally requires fewer parameters given the relatively few weights of each radial basis function. This feature can speed up training and is particularly useful when processing relatively small datasets.

[0077] Therefore, in this embodiment, in the traditional HQCGAN, the classic fully connected (FC) neural network is replaced by RBFNN. In this enhancement, the completed preprocessed sample image data is used as the input of the discriminator (denoted as D) in the form of a binary array. The discriminator uses the excellent nonlinear data processing capabilities of RBFNN to effectively classify and discriminate the model-generated data from the quantum generator, which is sampled by the measured data. Ideally, D cannot distinguish whether the input is from the real distribution or the generated model distribution. The final state of the generator (denoted as G) is then measured to align with the handwritten digital image of the real sample. This alignment is accomplished by training using a binary cross entropy loss function.

[0078] like Figure 4 As shown in Figure 1, the variational quantum circuit includes a rotation layer and an entanglement layer; the rotation layer uses a single qubit Ry gate. In addition, the introduction of quantum entanglement properties enhances the correlation between information, thereby generating patterns that are unattainable in classical systems. The rotation layer integrates single-qubit gates such as R x , R y and R z The gate facilitates the rotation of the qubit to manipulate its state. The entanglement layer uses rotation gates controlled by two qubits, especially CZ gates, to establish entanglement between qubits. Specifically, in the CZ gate, when the control qubit is set to 0, the operation on the target qubit is not affected. However, when the control qubit is 1, it triggers the corresponding operation on the target qubit, thereby inducing entanglement between the qubits involved in the gate operation. This structured circuit design approach enables the RBFQGAN algorithm to effectively utilize quantum principles and enhance its ability to generate and process quantum data. In this embodiment, CZ gates (two qubit gates) are used in the entanglement layer and connected in a ring topology (circle) to achieve low cost and stronger expressiveness and entanglement capabilities.

[0079] S4, adversarial training: the generated probability distribution learned by the generator and the preprocessed discrete sample data are input into the discriminator, and the features of the samples are extracted through the nonlinear data learning and processing capabilities of the radial basis network to determine the authenticity of the input image data samples;

[0080] S5, parameter optimization: training through the cross entropy loss function and optimizing the parameters through the Adam optimizer, and then passing the optimized parameters to the generator and discriminator for further training until the training losses of the generator and discriminator converge to the Nash equilibrium;

[0081] The cross entropy loss function is:

[0082]

[0083] Where N is the number of samples, y is the true label of the sample (0 or 1), and p is the probability predicted by the discriminator that the sample belongs to class 1.

[0084] S6, when the optimization training is completed, the image data samples generated by the output generator are output, that is, the image data samples generated by the RBF-QGAN training with the smallest distribution difference from the real image data.

[0085] Based on similar inventive concepts, an embodiment of the present invention further provides a computer storage medium storing a readable program, which can execute the above-mentioned image processing method based on RBF-QGAN when the program is running.

[0086] Based on similar inventive concepts, an embodiment of the present invention provides an electronic device, comprising: a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other through the communication bus;

[0087] The memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform an operation corresponding to the above-mentioned image processing method based on RBF-QGAN.

[0088] Based on similar inventive concepts, an embodiment of the present invention further provides a computer program product, including computer instructions, which instruct a computing device to perform operations corresponding to the above-mentioned image processing method based on RBF-QGAN.

[0089] Example 2

[0090] In this embodiment, the performance of RBF-QCAN is verified and evaluated;

[0091] In order to evaluate the stability of RBF-QGAN, the coefficient of variation (CV) is used as a key indicator for evaluating the stability of the discriminator performance in this embodiment. CV is a standardized indicator that compares the degree of dispersion of the data by calculating the ratio of the standard deviation to the mean of the original data set. The calculation of CV is a unitless scale that helps to compare the differences between different data sets or models. By using this indicator, the fluctuation of the loss value can be closely monitored throughout the training phase of the model, thereby gaining a detailed understanding of the overall training stability of the model components. The formula for calculating CV is as follows:

[0092]

[0093] Among them, S D and M D They are the standard deviation and mean of the loss value of the discriminator respectively; the smaller the CV value, the more stable the system.

[0094] Specifically in the experiment, a specific experimental setup was designed to more deeply study the effectiveness and stability of the RBF discriminator within the model training framework. In the training scheme of each model variant, the loss value associated with the discriminator was carefully recorded at each iteration. These collected data points can be used to calculate the CV of the loss value corresponding to each training method, thereby providing a quantifiable stability measure, such as Figure 5 shown.

[0095] Figure 5 (a)-(c) in the figure correspond to the CV values ​​of various network models on the MNIST dataset with MNIST Gaussian noise, MNIST uniform noise, and MNIST salt and pepper noise, respectively; it can be seen that the CV values ​​of QGAN trained on the MNIST dataset under three different noise conditions are affected by three different noises: Gaussian noise, uniform noise, and salt and pepper noise. This comparative study allows a comprehensive evaluation of the robustness and adaptability of our model under various noise interference conditions. Our results clearly highlight the effectiveness of our proposed RBF-QGAN, as evidenced by the consistently lower CV values ​​for all noise types compared to alternative QGAN architectures. This significant difference in CV values ​​emphasizes the robustness and stability of our model in recognizing and processing noisy image inputs. Enhanced stability is crucial in real-world applications, where noise is an inevitable aspect of data acquisition and processing. Therefore, our approach not only demonstrates excellent performance under noisy conditions, but also shows its potential for deployment in practical environments that require flexibility and reliable image processing capabilities.

[0096] like Figure 6 As shown, when evaluating the basic ability of image generation, the number "0" in the MNIST dataset was selected as the test object. The model was used to generate grayscale images corresponding to these selected categories. The algorithm showed the ability to quickly converge the loss value to the Nash equilibrium point. It is worth noting that when the cross entropy loss value of the generator and the discriminator stabilizes at around 0.7, it indicates that convergence is achieved at this time. In addition, the display of the generated image is extremely similar to the selected sample picture. These findings show the excellent ability of the RBF-QGAN model in performing the grayscale image generation task of the MNIST dataset.

[0097] The robustness experiment steps are as follows: The number "0" in the MNIST dataset is selected as the image sample for model training. Various levels of noise are introduced to test the robustness of the model. The evaluation combines qualitative observation and quantitative analysis, with special emphasis on the cross entropy loss value as a performance evaluation indicator. Specifically, by carefully examining the performance of HQCGAN, the loss values ​​of three different discriminators under different degrees of noise are included.

[0098] like Figure 7 As shown in the figure, under certain intensities of Gaussian noise, uniform noise, and salt and pepper noise, the loss value shows minimal fluctuations despite the presence of noise, and RBF-QGAN shows greater robustness. These results highlight the model's superiority in achieving similar results in the presence of fluctuating noise levels, thus reaffirming its robustness under adverse conditions.

[0099] In summary, the present invention utilizes the classical machine learning component RBFNN and significantly enhances its ability to generate high-fidelity grayscale images with discrete value distribution. Through meticulous experiments, evaluating the training cross-validation scores and the robustness of the loss function, the excellent performance of the QGAN model is demonstrated, especially in the presence of noisy input data. The results of this study highlight the importance of integrating classical machine learning techniques into the quantum framework to enhance the efficacy of quantum generative models. By incorporating RBFNN into quantum-inspired algorithms, novel perspectives are provided to address common challenges related to image quality and training stability. This marks a substantial progress in the development of quantum generative adversarial networks, positioning them as a viable solution for a wide range of practical applications.

[0100] Example 3

[0101] Based on the RBF-QGAN-based image processing method proposed in Example 1, in this embodiment, an RBF-QGAN-based image processing system is proposed, which specifically includes:

[0102] Data acquisition module: acquire image data;

[0103] Data preprocessing module: preprocess the acquired image data to obtain discrete sample data;

[0104] Model building module: Build an RBF-QGAN including a discriminator and a generator. The discriminator uses a radial basis network and the generator uses a variational quantum circuit. The generator receives discrete samples of input image data, learns the input generation probability distribution, and uses quantum circuits to generate output probability distributions. The discriminator receives the data output by the generator and determines the authenticity of the data.

[0105] Adversarial training module: The generated probability distribution learned by the generator and the preprocessed discrete sample data are input into the discriminator, and the features of the samples are extracted through the radial basis network to determine the authenticity of the data samples;

[0106] Parameter optimization module: training through the cross entropy loss function and parameter optimization through the Adam optimizer, and then passing the optimized parameters to the generator and discriminator for further training until the training losses of the generator and discriminator converge to the Nash equilibrium;

[0107] And, data sample generation module: after the optimization training is completed, the image data samples generated by the output generator are output.

[0108] The method of the present invention may be implemented in hardware, firmware, or as software or computer code that may be stored in a recording medium (such as a CDROM, RAM, floppy disk, hard disk, or magneto-optical disk), or as computer code that is originally stored in a remote recording medium or a non-temporary machine-readable medium downloaded over a network and will be stored in a local recording medium, so that the method described herein may be stored in such software processing on a recording medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware (such as an ASIC or FPGA). It is understood that a computer, processor, microprocessor controller, or programmable hardware includes a storage component (e.g., RAM, ROM, flash memory, etc.) that can store or receive software or computer code, and when the software or computer code is accessed and executed by a computer, processor, or hardware, the method described herein is implemented. In addition, when a general-purpose computer accesses the code for implementing the method shown herein, the execution of the code converts the general-purpose computer into a dedicated computer for executing the method shown herein.

[0109] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements all fall within the scope of the present invention to be protected.

Claims

1. An image processing method based on RBF-QGAN, characterized in that: The following steps are involved: Get image data; Preprocessing the acquired image data to obtain discrete sample data; Construct an RBF-QGAN consisting of a discriminator and a generator, where the discriminator uses a radial basis network and the generator uses a variational quantum circuit; The generator receives discrete samples, learns the input generation probability distribution, and uses quantum circuits to generate output probability distributions; the discriminator receives the data output by the generator and determines the authenticity of the data; The generated probability distribution learned by the generator and the preprocessed discrete sample data are input into the discriminator, and the features of the samples are extracted through the radial basis network to determine the authenticity of the data samples; The cross entropy loss function is used for training and the Adam optimizer is used for parameter optimization. The optimized parameters are then passed into the generator and discriminator for further training until the training losses of the generator and discriminator converge to the Nash equilibrium. After the optimization training is completed, the image data samples generated by the generator are output.

2. The image processing method based on RBF-QGAN according to claim 1, characterized in that: The steps of image data preprocessing include: Reduce the image data to image data with a dimension of 16×16; The image data after dimensionality reduction is blurred and normalized; The processed image data is converted into discrete sample data as the true distribution.

3. The image processing method based on RBF-QGAN according to claim 1, characterized in that: The radial basis network includes: an input layer, a hidden layer and an output layer; the radial basis network uses a radial basis function as a Gaussian kernel function, which is expressed as: H(x,c)=e(-β·||x-c|| 2 ) where ||·|| is the norm on the function space, x i represents the i-th sample in the training set, x is the input data, c is the center of the RBF layer, and β is defined as the expansion coefficient.

4. The image processing method based on RBF-QGAN according to claim 3, characterized in that: The output y of the radial basis network i for: Where ω0 represents the bias of the output neuron, q represents the number of hidden neurons, and w j represents the weight between the jth hidden neuron and the output neuron; v j and σ j They represent the center and width of the jth hidden neuron respectively; g1 represents the Gaussian kernel function.

5. The image processing method based on RBF-QGAN according to claim 1, characterized in that: The variational quantum circuit includes a rotation layer and an entanglement layer; the rotation layer uses a single quantum bit Ry gate; the entanglement layer uses a CZ gate and is connected in a ring topology.

6. An image processing system based on RBF-QGAN, characterized in that: include: Data acquisition module: acquire image data; Data preprocessing module: preprocess the acquired image data to obtain discrete sample data; Model building module: Build an RBF-QGAN including a discriminator and a generator. The discriminator uses a radial basis network and the generator uses a variational quantum circuit. The generator receives discrete samples, learns the input generation probability distribution, and uses quantum circuits to generate output probability distributions. The discriminator receives the data output by the generator and determines the authenticity of the data. Adversarial training module: The generated probability distribution learned by the generator and the preprocessed discrete sample data are input into the discriminator, and the features of the samples are extracted through the radial basis network to determine the authenticity of the data samples; Parameter optimization module: trains through the cross entropy loss function and optimizes parameters through the Adam optimizer, then passes the optimized parameters to the generator and discriminator for further training until the training losses of the generator and discriminator converge to the Nash equilibrium; And, data sample generation module: after the optimization training is completed, the image data samples generated by the output generator are output.

7. A computer storage medium storing a readable program, characterized in that: When the program is running, it can execute the image processing method based on RBF-QGAN described in any one of claims 1 to 5.

8. An electronic device, characterized in that: include: A processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other via the communication bus; The memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform an operation corresponding to an image processing method based on RBF-QGAN according to any one of claims 1 to 5.

9. A computer program product comprising computer instructions, characterized in that The computer instructions instruct the computing device to perform operations corresponding to the image processing method based on RBF-QGAN as described in any one of claims 1 to 5.

10. A hybrid quantum generative adversarial network based on radial basis network, characterized in that: It includes a discriminator and a generator. The discriminator adopts a radial basis network. The generator receives discrete samples of input image data, learns the input generation probability distribution, and uses a quantum circuit to generate an output probability distribution. The discriminator receives the data output by the generator and judges the authenticity of the data.

Images