Hyperspectral image defogging method based on quantum generative adversarial network

Abstract

The invention provides a hyperspectral image defogging method based on a quantum generative adversarial network, and the method comprises the following steps: 10, obtaining the quantum generative adversarial network through employing a training data set; the quantum generative adversarial network comprises a quantum generator and a discriminator; and step 20, defogging the to-be-processed foggy hyperspectral image by using the quantum generative adversarial network. According to the hyperspectral image defogging method based on the quantum generative adversarial network provided by the invention, the image defogging precision and efficiency are improved, and the use of quantum bits is reduced.

Description

Technical Field

[0001] The present invention relates to the field of quantum machine learning technology, and in particular to a hyperspectral image dehazing method based on a quantum generative adversarial network. Background Art

[0002] Hyperspectral images are widely used in remote sensing, environmental monitoring and agriculture due to their rich spectral information. However, under complex weather conditions (such as haze environment), hyperspectral images are often affected by spectral distortion and reduced spatial resolution, which affects the accuracy of subsequent analysis tasks. At the same time, the impact of haze pollution on different spectral wavelengths is different. Bands with shorter wavelengths are usually more sensitive to haze, and conversely, bands with longer wavelengths are more robust to haze. Existing methods mainly rely on classical machine learning or deep learning frameworks. Although they can achieve certain results, hyperspectral images contain rich spectral information, high data dimensions, and complex correlations between spectra. It is difficult for traditional methods to effectively capture these correlations.

[0003] At present, generative adversarial networks are used for dehazing hyperspectral images. Generative adversarial networks are composed of two neural networks, namely the generator and the discriminator. Through the adversarial game between the generator and the discriminator, high-quality samples similar to the real data distribution are generated. However, the classic generative adversarial network requires a lot of computing resources to process high-dimensional data, and is less efficient for tasks such as hyperspectral images. At the same time, the classic generative adversarial network needs a large number of parameters and deep network layers to fit the complex distribution of high-dimensional data, which is prone to problems such as overfitting and unstable training. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a hyperspectral image defogging method based on a quantum generative adversarial network, improve the accuracy and efficiency of image defogging, and reduce the use of quantum bits.

[0005] In order to solve the above technical problems, the present invention provides a hyperspectral image dehazing method based on quantum generative adversarial network, comprising the following steps: Step 10, using the training data set to obtain a quantum generative adversarial network; the quantum generative adversarial network includes a quantum generator and a discriminator; Step 20, using a quantum generative adversarial network to dehaze the foggy hyperspectral image to be processed.

[0006] As a further improvement of the present invention, the step 10 specifically includes: Step 101, constructing a training data set, wherein the training data set includes a plurality of image pairs consisting of a hazy hyperspectral image and a corresponding hazy hyperspectral image; Step 102, quantum encoding the foggy hyperspectral image using a band register and a pixel register; Step 103, inputting the quantum-encoded waveband into a quantum generator; Step 104, measuring the result of the quantum generator, obtaining pixel information of all bands of the quantum generator, and reconstructing a defogged hyperspectral image; Step 105, using the defogged hyperspectral image and the corresponding non-fog hyperspectral image in the training data set, the quantum generator and the discriminator are trained to obtain a trained quantum generative adversarial network.

[0007] As a further improvement of the present invention, step 102 specifically includes: The band register uses amplitude coding to encode the band information of the foggy hyperspectral image; The pixel register encodes the pixel information of the foggy hyperspectral image through angle encoding, and binds the band information and the corresponding pixel information into a joint quantum state.

[0008] As a further improvement of the present invention, the pixel register encodes the pixel information of the foggy hyperspectral image by angle encoding, and binds the band information and the corresponding pixel information to a joint quantum state, specifically including: The pixel register encodes the pixel value of the foggy hyperspectral image into the quantum state using angle encoding. The controlled rotating gate is used to control the rotation angle of the pixel register according to the state of the band register, so that each pixel value is mapped to the rotation angle of the quantum bit, realizing the encoding of the pixel value, thereby binding the band index and the corresponding pixel value to the joint quantum state.

[0009] As a further improvement of the present invention, the band register uses quantum bits to encode the band index, specifically including: A Hadamard gate is applied to the band register to generate a uniform superposition state to represent all band indices.

[0010] As a further improvement of the present invention, in step 102, before performing quantum coding, the multiple bands of the foggy hyperspectral image are first grouped to form multiple band groups; then the multiple band groups are quantum coded using the band register and the pixel register respectively; In step 103, a corresponding quantum sub-generator is established for each band group, and the quantum-encoded band groups are respectively input into the corresponding quantum sub-generators; In step 104, the results of each quantum sub-generator are measured respectively, the pixel information of all bands of each quantum sub-generator is obtained, and the defogging hyperspectral image is reconstructed.

[0011] As a further improvement of the present invention, step 104 specifically includes: The measurement results of the quantum generator are obtained through multiple Pauli-Z measurements; the expected value of the pixel bit measurement of each band index is counted, and the expected value of the measurement of each quantum bit is mapped to the pixel value to obtain the pixel information of all bands of the quantum generator; the pixel information of all bands is spliced ​​to obtain the dehazed hyperspectral image.

[0012] As a further improvement of the present invention, step 105 specifically includes: The dehazed hyperspectral image and the corresponding haze-free hyperspectral image in the training data set are input into the discriminator, and the quantum generator and discriminator are alternately gradient optimized according to the discrimination results and the loss function until the quantum generative adversarial network reaches the convergence state of Nash equilibrium, thereby obtaining the trained quantum generative adversarial network.

[0013] As a further improvement of the present invention, the loss function includes the loss of the generator and the loss of the discriminator; the loss of the generator includes an adversarial loss term and a content loss term, the adversarial loss term is used to measure the ability of the quantum generator to deceive the discriminator, and the content loss term is used to measure the similarity between the dehazed hyperspectral image and the haze-free hyperspectral image, including the minimum mean square error, inter-spectral gradient loss and spectral angle constraint; the loss of the discriminator is the cross entropy loss in the quantum generative adversarial network.

[0014] As a further improvement of the present invention, the step 20 specifically includes: Step 201, grouping multiple bands of the foggy hyperspectral image to be processed to form multiple band groups; Step 202, quantum encoding each band group using a band register and a pixel register; Step 203, inputting the quantum-encoded band groups into the corresponding quantum sub-generators in the quantum generative adversarial network respectively; Step 204, measuring the result of each quantum sub-generator, obtaining the pixel information of all bands of each quantum sub-generator, and reconstructing the defogging hyperspectral image.

[0015] The invention provides a method for defogging hyperspectral images based on quantum generative adversarial networks, which utilizes quantum superposition and quantum entangled states to model high-dimensional space, reduces the dependence on parameter quantities, has higher representation capabilities, can more effectively capture the characteristics of hyperspectral data, and is suitable for high-dimensional feature representation of hyperspectral data. At the same time, quantum computing has powerful parallel computing capabilities, allowing high-dimensional data to be processed quickly in exponential state spaces, so that quantum generators have significant efficiency advantages in modeling complex distributions between hyperspectral bands, and improve the accuracy and efficiency of image defogging.

[0016] The method of the present invention utilizes the strong correlation between multiple bands in the hyperspectral image, especially the characteristic that adjacent bands usually reflect similar spectral information. By grouping the bands of the hyperspectral image according to the spectral correlation, the redundant information is effectively reduced while retaining the complementary characteristics between the bands. A sub-generator is assigned to each group so that each sub-generator focuses on processing a group of related bands. The sub-generator can be specifically optimized within a specific band group to improve the defogging accuracy. The training task of each sub-generator becomes more simplified and focused, which can not only reduce interference but also improve the computational efficiency in the defogging process.

[0017] The method of the present invention uses quantum superposition to encode band information instead of encoding pixel information of all bands one by one, which improves the efficiency of data processing and can reduce the complexity of quantum circuits. Thereby reducing the computational cost of training and reasoning, so that efficient image processing can still be achieved in a hardware-constrained environment. This method not only improves computing efficiency, but also reduces resource consumption, demonstrating the advantages of quantum computing in the field of image processing. In the quantum generative adversarial network, the quantum circuit optimizes and processes the spectral information between bands and the spatial information of pixels in a unified quantum state space. This coupled modeling helps to simultaneously restore the spatial and spectral characteristics of the image and improve the dehazing accuracy.

[0018] The parameterized quantum circuit of the method of the present invention can efficiently extract the features of hyperspectral images through the combination of quantum gates. By introducing adjustable parameters (such as the rotation angle of the quantum gate), the quantum circuit can flexibly adapt to different task requirements. In the scenario of hyperspectral image dehazing, the quantum generator and discriminator in QGAN can accurately control the quality of the generated image and the dehazing effect by adjusting these quantum parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 A flowchart of a hyperspectral image dehazing method based on a quantum generative adversarial network provided in an embodiment of the present invention; Figure 2 Schematic diagram of the structure of a quantum coding module in an embodiment of the present invention; Figure 3 4 is a circuit structure diagram of a quantum generator in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings.

[0021] Quantum Generative Adversarial Network (QGAN) is a generative adversarial network based on quantum computing. It uses the advantages of quantum superposition and parallelism to generate complex data distributions or solve high-dimensional problems. It aims to solve the bottleneck problem of classical generative adversarial networks relying on computing power, bringing potential performance improvements to generative models and adversarial models.

[0022] This paper proposes to use QGAN for the hyperspectral image defogging task and provides a hyperspectral image defogging method based on quantum generative adversarial network. Figure 1 As shown, the following steps are included: Step 10: Using the training data set, a quantum generative adversarial network is obtained. The quantum generative adversarial network includes a quantum generator and a discriminator.

[0023] Step 20, using a quantum generative adversarial network to dehaze the foggy hyperspectral image to be processed.

[0024] The present invention utilizes quantum generative adversarial networks to perform dehazing of hyperspectral images, which can effectively utilize spectral and spatial information, thereby improving the accuracy and efficiency of image dehazing.

[0025] As a preferred example, step 10 specifically includes: Step 101 : construct a training data set, where the training data set includes a plurality of image pairs consisting of one-to-one corresponding hyperspectral images with fog and hyperspectral images without fog.

[0026] Specifically, data augmentation operations are performed on all hyperspectral images, including resizing, center cropping, conversion to tensors, and normalization.

[0027] Step 102: quantum encode the foggy hyperspectral image using a band register and a pixel register.

[0028] In quantum coding, band registers and pixel registers are used to encode band index and pixel value for each band of the foggy hyperspectral image, respectively.

[0029] Preferably, step 102 specifically includes: The band register uses amplitude coding to encode the band index of the foggy hyperspectral image.

[0030] Specifically, a Hadamard gate is applied to the band register to generate a uniform superposition state to represent all band indexes.

[0031] The pixel register encodes the pixel information of the foggy hyperspectral image through angle encoding, and binds the band information and the corresponding pixel information into a joint quantum state.

[0032] Specifically, the pixel register uses angle encoding to encode the pixel values ​​of the foggy hyperspectral image into the quantum state; using a controlled rotating gate, the rotation angle of the pixel register is controlled according to the state of the band register, so that each pixel value is mapped to the rotation angle of the quantum bit, realizing the encoding of the pixel value, thereby binding the band index and the corresponding pixel value to the joint quantum state.

[0033] For example, the size of a foggy hyperspectral image is N×N×M, the pixel size is N×N, and the number of bands is M.

[0034] like Figure 2 As shown, use qubits encode M band indices, through A Hadamard gate is used to generate a uniform superposition state to represent all bands: Formula (1) In the formula, The quantum state representing the band index covers the superposition state of all M bands. Indicates the status of the bth band.

[0035] use qubits, initialize the pixel register to all zeros: Formula (2) In the formula, Indicates the initial state of the pixel register. At initialization, all bits are set to all zeros. , these qubits will be updated to quantum states reflecting pixel values ​​by subsequent encoding operations.

[0036] Flatten the N×N two-dimensional pixel matrix into a one-dimensional vector, normalize the pixel value of each pixel, and the rotation angle corresponding to each normalized pixel is: Formula (3) In the formula, represents the rotation angle of the kth pixel, represents the pixel value of the kth pixel, ∈[0,1].

[0037] According to the state of the band information specified by the band register, a CRY gate is applied to each quantum bit of the pixel register using a controlled rotating gate, thereby encoding the pixel value of each band as a quantum state. , the pixel value Mapped to the amplitude of the quantum bit, the resulting quantum state can be expressed as follows: Formula (4) The function of the controlled rotation gate is to control the rotation of the quantum bit in the pixel register according to the state of the band index, where Represents the quantum state corresponding to the pixel value in each band after the rotation operation.

[0038] The quantum state of all pixel values ​​in a specific band after encoding can be expressed as: Formula (5) In the formula, represents the quantum state of all pixel values ​​in the b-th band, Indicates the rotation angle of the k-th pixel in the b-th band.

[0039] By establishing entanglement between the band register and the pixel register, the band index in the band register and the pixel information in the pixel register are bound. The hyperspectral image can be represented by the joint quantum state of the band register and the pixel register: Formula (6) In the formula, Represents the joint quantum state of the entire hyperspectral image after quantum encoding.

[0040] Existing quantum coding methods for generating hyperspectral images usually treat all bands as independent inputs, requiring the allocation of a large number of quantum bits to encode the pixel information of all bands one by one. This coding strategy is easily constrained by hardware resource limitations when processing large-scale hyperspectral data, thereby limiting the practical application effect of quantum algorithms and not fully utilizing the correlation between bands.

[0041] In step 102 of the embodiment of the present invention, quantum coding utilizes quantum superposition to efficiently compress information of multiple bands into fewer quantum bits, thereby reducing the consumption of quantum hardware resources. At the same time, by establishing entanglement between the band register and the pixel register, the band index in the band register and the pixel information in the pixel register are bound, and the quantum bit can simultaneously represent the pixel values ​​and spatial positions of multiple bands, and efficient coupling can be achieved between the band register and the pixel register. Based on the parallel processing of quantum circuits, quantum circuits can process all pixels of a certain band in parallel in one operation without rotating each pixel one by one.

[0042] Step 103, input the quantum-encoded band into the quantum generator.

[0043] The quantum generator includes several quantum rotating gates RX with parameters, several quantum rotating gates RY with parameters and several controlled NOT gates, and several corresponding circuits are arranged in parallel in sequence, such as Figure 3 shown.

[0044] The qubits encoding pixel information in the pixel register are passed through a parameterized rotation gate RY(θ), and then entanglement is introduced between the qubits through a parameterized XX gate. The XX gate is a two-qubit gate that introduces entanglement between two qubits through a rotation operation. The entanglement layer shares local information to the entire quantum state, allowing the global generator state to capture pixel correlations.

[0045] On each qubit, a single-qubit rotation gate RX(θ) is applied to adjust the phase and amplitude of a single qubit to improve the expressiveness of the generator. Next, an XX gate is used on each adjacent qubit to establish the correlation between the characteristic qubits, so that the circuit can capture the interdependence between pixel registers. In the last layer, each qubit adjusts the probability distribution of the quantum state output through a single-qubit rotation gate RY(θ).

[0046] Step 104, measuring the result of the quantum generator, obtaining pixel information of all bands of the quantum generator, and reconstructing the dehazed hyperspectral image.

[0047] Preferably, the measurement results of the quantum generator are obtained by multiple Pauli-Z measurements. The expected value of the pixel bit measurement of each band index is counted, and the expected value of the measurement of each quantum bit is mapped to the pixel value to obtain the pixel information of all bands of the quantum generator. The pixel values ​​of all bands are spliced ​​to obtain the defogging hyperspectral image.

[0048] The measurement operation is to restore the quantum state to the classical state, specifically: Measurement Band Register qubits, obtain the band index b, and the measurement of the band register will collapse the joint quantum state to the substate corresponding to the specific band b: Formula (7) At this time, the state of the pixel register collapses to the pixel quantum state corresponding to the index of band b: Formula (8) Measuring pixel register N 2 qubits, calculate the expected value of each pixel qubit: Formula (9) In the formula, V represents the measurement operator, represents the quantum state to be measured, and represents the measurement operator V in the quantum state The expected value below.

[0049] Quantum state After multiple measurements, the average result for the measurement operator V is calculated.

[0050] Specifically, a set of measurement operators V are defined to measure the state The measurement basis is a set of orthogonal quantum states, usually expressed as When a quantum state is measured, the measurement result will correspond to these ground states, and the quantum state Measure and obtain the measurement basis The probability is: After multiple measurements, the probability distribution of a certain ground state will be obtained, and the expected value of the measurement operator V will be calculated.

[0051] The expected value of each quantum bit is mapped to a pixel value. After several measurements, the pixel value of each band index b is obtained. The pixel values ​​of all bands b=0, 1, …, M−1 are concatenated to reconstruct the dehazed hyperspectral image.

[0052] Step 105, using the defogged hyperspectral image and the corresponding non-fog hyperspectral image in the training data set, the quantum generator and the discriminator are trained to obtain a trained quantum generative adversarial network.

[0053] In this embodiment, the discriminator adopts a classical neural network, whose input is the defogged hyperspectral image reconstructed by the quantum generator and the real haze-free hyperspectral image. A score is output through the last layer of Sigmoid activation function to evaluate the difference between the defogged image and the real haze-free image. The closer the output is to 1, the closer the defogged image is to the real haze-free image.

[0054] Specifically include: The dehazed hyperspectral image and the corresponding haze-free hyperspectral image in the training data set are input into the discriminator, and the quantum generator and discriminator are alternately gradient optimized according to the discrimination results and the loss function until the quantum generative adversarial network reaches the convergence state of Nash equilibrium, thereby obtaining the trained quantum generative adversarial network.

[0055] Preferably, the loss function includes the loss of the generator and the loss of the discriminator.

[0056] The loss of the quantum generator includes adversarial loss and content loss. The adversarial loss is used to measure the ability of the quantum generator to deceive the discriminator. The expression is: Formula (10) In the formula, represents the adversarial loss of the quantum generator, represents the output of the quantum discriminator, represents the image generated by the generator, represents the set of parameters of the quantum discriminator.

[0057] The content loss term is used to measure the similarity between the dehazed hyperspectral image and the haze-free hyperspectral image, including the minimum mean square error, inter-spectral gradient loss, and spectral angle constraint.

[0058] The minimum mean square error is used to calculate the difference in pixels between the dehazed hyperspectral image and the non-hazed hyperspectral image. The expression is: Formula (11) In the formula, represents the minimum mean square error loss, represents the height of the hyperspectral image, represents the width of the hyperspectral image, Indicates the number of spectral bands of the hyperspectral image, represents the pixel value of the haze-free hyperspectral image at pixel position (h, w) and band c, Represents the pixel value of the dehazed hyperspectral image at pixel position (h, w) and band c.

[0059] The inter-spectral gradient term refers to the gradient difference between adjacent bands in the spectral curve, which can be used to constrain the smoothness of the spectral curve. The expression is: Formula (12) In the formula, represents the inter-spectral gradient loss, Represents band gradients: .

[0060] The spectral angle constraint calculates the angle difference between the spectral vectors of each pixel to measure the spectral similarity, and the expression is: Formula (13) In the formula, represents the spectral angle constraint loss, represents the inner product of the spectral vector, Represents the L2 norm of the spectral vector.

[0061] The loss of the discriminator is the cross entropy loss in the quantum generative adversarial network, expressed as: Formula (14) In the formula, represents the loss function of the discriminator.

[0062] Preferably, the process of training the quantum generative adversarial network is: fix the quantum generator parameter θ, optimize the discriminator parameter φ, and minimize the discriminator loss ;Fix the discriminator parameter φ, optimize the quantum generator parameter θ, and minimize the loss of the quantum generator The quantum generator and the discriminator are optimized alternately until convergence. The quantum generator gradually improves the quality of the dehazed image and eventually generates a dehazed hyperspectral image that is close to the real haze-free image.

[0063] As a preferred example, in step 102, before performing quantum encoding, multiple bands of the foggy hyperspectral image are first grouped to form multiple band groups.

[0064] Specifically, the spectral correlation coefficient between bands is calculated using formula (15): Formula (15) In the formula, represents the correlation coefficient between the i-th band and the j-th band, represents the covariance of the spectral reflectance of the i-th band and the spectral reflectance of the j-th band, represents the variance of the spectral reflectance of the i-th band, represents the variance of the spectral reflectance of the jth band, represents the spectral reflectance of the i-th band, represents the spectral reflectance of the j-th band.

[0065] The bands with higher spectral correlation coefficients are divided into the same band group, and the haze pollution levels of the bands in the same band group are similar.

[0066] Then, the band registers and pixel registers are used to perform quantum encoding on multiple band groups respectively.

[0067] In step 103, a corresponding quantum sub-generator is established for each band group, and the quantum-encoded band groups are respectively input into the corresponding quantum sub-generators.

[0068] In step 104, the results of each quantum sub-generator are measured respectively, the pixel information of all bands of each quantum sub-generator is obtained, and the defogging hyperspectral image is reconstructed.

[0069] In step 105, for multiple quantum sub-generators, the parameters of the quantum generator include the parameters of all quantum sub-generators, θ=[ ]. By jointly updating and optimizing the parameters of all quantum sub-generators, each quantum sub-generator can not only focus on the generation of a local band group, but also collaboratively improve the quality of the overall image generation.

[0070] In the quantum circuit of the quantum generator, the gradient of the parameterized quantum gate is calculated by the parameter shift rule. The gradient is passed to the loss function through back propagation to update the quantum generator parameters. The gradient of the jth parameter of the i-th quantum sub-generator with respect to the loss is: Formula (16) In the formula, represents the loss function of the quantum generator, Represents the jth parameter of the i-th quantum sub-generator in the quantum sub-generator.

[0071] Different bands may be affected by haze to varying degrees. Based on the correlation between spectral bands, an embodiment of the present invention uses a clustering algorithm to divide the hyperspectral image into multiple spectral subsets, and constructs an independent quantum sub-generator for each subset. Each quantum sub-generator is trained separately, so that each sub-generator focuses on dehazing a specific band, achieving difference processing between bands, thereby improving the details and realism of the generated image.

[0072] Step 20 specifically includes: Step 201 : grouping multiple bands of the foggy hyperspectral image to be processed to form multiple band groups.

[0073] Step 202: quantum encode each band group using a band register and a pixel register.

[0074] Step 203, input the quantum-encoded band groups into the corresponding quantum sub-generators in the quantum generative adversarial network respectively.

[0075] Step 204, measuring the result of each quantum sub-generator, obtaining the pixel information of all bands of each quantum sub-generator, and reconstructing the dehazed hyperspectral image.

[0076] 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 specific embodiments. The above specific embodiments and the description in the specification are only for further illustrating 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, which fall within the scope of the present invention to be protected. The scope of the present invention to be protected is defined by the claims and their equivalents.

Claims

1. A hyperspectral image dehazing method based on quantum generative adversarial network, characterized in that: The following steps are involved: Step 10, using the training data set to obtain a quantum generative adversarial network; The quantum generative adversarial network includes a quantum generator and a discriminator; Step 20, using a quantum generative adversarial network to dehaze the foggy hyperspectral image to be processed.

2. The hyperspectral image dehazing method based on quantum generative adversarial network according to claim 1 is characterized in that: The step 10 specifically includes: Step 101, constructing a training data set, wherein the training data set includes a plurality of image pairs consisting of a hazy hyperspectral image and a corresponding hazy hyperspectral image; Step 102, quantum encoding the foggy hyperspectral image using a band register and a pixel register; Step 103, inputting the quantum-encoded waveband into a quantum generator; Step 104, measuring the result of the quantum generator, obtaining pixel information of all bands of the quantum generator, and reconstructing a defogged hyperspectral image; Step 105, using the defogged hyperspectral image and the corresponding non-fog hyperspectral image in the training data set, the quantum generator and the discriminator are trained to obtain a trained quantum generative adversarial network.

3. The hyperspectral image dehazing method based on quantum generative adversarial network according to claim 2 is characterized in that: The step 102 specifically includes: The band register uses amplitude coding to encode the band information of the foggy hyperspectral image; The pixel register encodes the pixel information of the foggy hyperspectral image through angle encoding, and binds the band information and the corresponding pixel information into a joint quantum state.

4. The hyperspectral image dehazing method based on quantum generative adversarial network according to claim 3 is characterized in that: The pixel register encodes the pixel information of the foggy hyperspectral image by angle encoding, and binds the band information and the corresponding pixel information to a joint quantum state, specifically including: The pixel register encodes the pixel value of the foggy hyperspectral image into the quantum state using angle encoding. The controlled rotating gate is used to control the rotation angle of the pixel register according to the state of the band register, so that each pixel value is mapped to the rotation angle of the quantum bit, realizing the encoding of the pixel value, thereby binding the band index and the corresponding pixel value to the joint quantum state.

5. The hyperspectral image dehazing method based on quantum generative adversarial network according to claim 3 is characterized in that: The band register uses quantum bits to encode the band index, specifically including: A Hadamard gate is applied to the band register to generate a uniform superposition state to represent all band indices.

6. The hyperspectral image dehazing method based on quantum generative adversarial network according to claim 2 is characterized in that: In step 102, before performing quantum coding, the multiple bands of the foggy hyperspectral image are first grouped to form multiple band groups; then the multiple band groups are quantum coded using the band register and the pixel register respectively; In step 103, a corresponding quantum sub-generator is established for each band group, and the quantum-encoded band groups are respectively input into the corresponding quantum sub-generators; In step 104, the results of each quantum sub-generator are measured respectively, the pixel information of all bands of each quantum sub-generator is obtained, and the defogging hyperspectral image is reconstructed.

7. The hyperspectral image dehazing method based on quantum generative adversarial network according to claim 2 is characterized in that: The step 104 specifically includes: The measurement results of the quantum generator are obtained through multiple Pauli-Z measurements; the expected value of the pixel bit measurement of each band index is counted, and the expected value of the measurement of each quantum bit is mapped to the pixel value to obtain the pixel information of all bands of the quantum generator; the pixel information of all bands is spliced ​​to obtain the dehazed hyperspectral image.

8. The hyperspectral image dehazing method based on quantum generative adversarial network according to claim 2 is characterized in that: The step 105 specifically includes: The dehazed hyperspectral image and the corresponding haze-free hyperspectral image in the training data set are input into the discriminator, and the quantum generator and discriminator are alternately gradient optimized according to the discrimination results and the loss function until the quantum generative adversarial network reaches the convergence state of Nash equilibrium, thereby obtaining the trained quantum generative adversarial network.

9. The hyperspectral image dehazing method based on quantum generative adversarial network according to claim 8 is characterized in that: The loss function includes the loss of the generator and the loss of the discriminator; the loss of the generator includes an adversarial loss term and a content loss term, the adversarial loss term is used to measure the ability of the quantum generator to deceive the discriminator, and the content loss term is used to measure the similarity between the dehazed hyperspectral image and the haze-free hyperspectral image, including the minimum mean square error, inter-spectral gradient loss and spectral angle constraint; the loss of the discriminator is the cross entropy loss in the quantum generative adversarial network.

10. The hyperspectral image dehazing method based on quantum generative adversarial network according to claim 6, characterized in that: The step 20 specifically includes: Step 201, grouping multiple bands of the foggy hyperspectral image to be processed to form multiple band groups; Step 202, quantum encoding each band group using a band register and a pixel register; Step 203, inputting the quantum-encoded band groups into the corresponding quantum sub-generators in the quantum generative adversarial network respectively; Step 204, measuring the result of each quantum sub-generator, obtaining the pixel information of all bands of each quantum sub-generator, and reconstructing the defogging hyperspectral image.

Images