Hyperspectral image reconstruction identification method based on quantum generative adversarial network and quantum neural network

Abstract

The invention provides a method for recognizing a target after reconstruction of a hyperspectral image by using a quantum generative adversarial network and a quantum neural network in quantum machine learning, and the method comprises the following steps: 10) carrying out the preprocessing of a to-be-measured hyperspectral image, and obtaining a standardized low-resolution hyperspectral image; step 20) inputting the preprocessed image into a quantum generative adversarial network for image reconstruction to obtain a hyperspectral reconstruction image; and step 30) inputting the reconstructed hyperspectral image into a target recognition model based on the quantum neural network, and outputting a repaired image and a recognition result. According to the hyperspectral image reconstruction identification method based on the quantum generative adversarial network and the quantum neural network, the hyperspectral image is not influenced by low resolution, and the identification efficiency and accuracy are improved.

Description

Technical Field

[0001] This invention relates to the fields of quantum machine learning and image recognition technology, and specifically designs a hyperspectral image recognition method based on quantum machine learning. Background Technology

[0002] Hyperspectral remote sensing is an advanced detection method capable of simultaneously acquiring spatial and image information as well as continuous spectral information, providing a wealth of data. In recent years, this technology has achieved significant results in environmental monitoring, geological exploration, and other fields, attracting widespread attention. Hyperspectral images contain information from dozens to hundreds of continuous spectral bands, forming a complete data cube that contains rich spatial and spectral feature information. How to process and analyze this data to obtain the corresponding results is a crucial issue in this field. Important techniques in this area include hyperspectral image inpainting and reconstruction, and image recognition. These techniques can enhance the application value of images, improve the accuracy of target identification and land cover classification, while ensuring the preservation of spectral feature information.

[0003] Traditional hyperspectral image inpainting and target recognition techniques are primarily based on classic deep learning models, such as SRGAN and CNN network structures. While these methods can improve image quality and recognition accuracy to some extent, they still have limitations, mainly in their limited ability to preserve spectral features. This is especially true for hyperspectral images, which involve computationally complex data with small sample sizes, leading to poor data adaptability. To address this, researchers like Ben pioneered the introduction of quantum computing into image processing, proposing a quantum generative adversarial network (GAN) approach for hyperspectral image data augmentation. Building upon this research, numerous subsequent findings have led to the development of more practical quantum image processing methods.

[0004] While research on quantum machine learning in hyperspectral image processing has yielded increasingly rich results, it also has significant shortcomings. Limited by the current state of quantum hardware development, it is difficult to implement effectively on medium-sized noisy quantum computers in the near future. Furthermore, with increasing quantum depth, the training stability of the entire model is severely affected by quantum noise. Based on this, an improvement to the structure was proposed, using a hybrid quantum-classical generative adversarial network architecture to better adapt to the current hardware environment. However, the following problems still arise: 1) The design of the quantum generator suffers from the vanishing gradient problem. As the number of quantum circuit layers increases, the model training is prone to a plateau phenomenon, leading to difficulties in parameter optimization; 2) If the discrimination capability of the quantum discriminator is too strong, it will result in excessive consumption of computational resources and instability of training results; 3) The existing architecture of quantum neural networks is prone to the loss of spectral information when processing multi-band features of the hyperspectral spectrum. Summary of the Invention

[0005] The technical problem to be solved by this invention is: a hyperspectral image reconstruction and recognition method based on quantum generative adversarial networks and quantum neural networks, which reconstructs low-resolution hyperspectral images into high-quality hyperspectral images, thereby improving the detection efficiency and accuracy of target recognition.

[0006] To solve the above-mentioned technical problems, the embodiments of the present invention adopt the following technical solutions:

[0007] This invention provides a hyperspectral image reconstruction and recognition method based on quantum generative adversarial networks and quantum neural networks, comprising the following steps:

[0008] A method for reconstructing and identifying targets in hyperspectral images based on quantum generative adversarial networks and quantum neural networks, characterized by comprising the following steps:

[0009] Step 10) Preprocess the hyperspectral image to be tested to obtain a standardized low-resolution hyperspectral image;

[0010] Step 20) Input the preprocessed image into a quantum generative adversarial network for image reconstruction to obtain a hyperspectral reconstructed image;

[0011] Step 30) Input the reconstructed hyperspectral image into the target recognition model based on quantum neural network, and output the repaired image and the recognition result.

[0012] As a further improvement to this embodiment of the invention, step 10) specifically includes:

[0013] Step 101) Crop the original input hyperspectral image to generate image blocks of uniform size;

[0014] Step 102) Perform Gaussian blur and downsampling on the cropped image to generate low-resolution training samples;

[0015] Step 103) Normalize the image data, mapping pixel values ​​to the range [-1, 1]. The normalization formula is as follows:

[0016]

[0017] In the formula, I represents the original image, and max(I) represents the maximum pixel value in the image. G Normalized image;

[0018] Step 104) The normalized hyperspectral image is converted into a quantum state through quantum encoding. The encoding function represents:

[0019]

[0020] In the formula, ψ in Input quantum state, I′ xy,z The normalized hyperspectral pixel value at position (x, y, z), γ z Let |x>|y)|z> be the weighting factor for the z-th spectral band, where |x>|y)|z> represent the quantum basis of the spatial dimension and the spectral dimension, respectively. The denominator is a normalization factor that ensures that the modulus of the quantum state is 1.

[0021] As a further improvement to this embodiment of the invention, the reconstruction model based on quantum generative adversarial networks is obtained through the following steps:

[0022] Step 011) Construct a generator for a quantum generative adversarial network, using a quantum residual network structure;

[0023] Step 012) Construct the discriminator of the quantum generative adversarial network, using a classic binary classification convolutional neural network structure.

[0024] As a further improvement to this embodiment of the invention, step 011) specifically includes:

[0025] Step 0111) Input the quantum state information of the hyperspectral image after quantum encoding obtained in step 104);

[0026] Step 0112) Construct a multi-layer quantum circuit composed of quantum gates. The quantum convolution operation is represented as follows:

[0027]

[0028] In the formula, L gen The number of circuit layers in the quantum generator. Here are the training parameters for the i-th qubit in the l-th layer, where n is the total number of qubits, and CNOT is the training parameter for the i-th qubit. k,k+1 This is a controlled NOT gate operation with the control bit at k and the target bit at k+1;

[0029] Step 0113) Construct a quantum residual network and establish skip connections between convolutional layers. The connection mechanism is expressed by the following formula:

[0030] |ψ out >=αU(θ)|ψ in >+β|ψ in > Equation (4)

[0031] In the formula, α and β are the weight coefficients of the learnable quantum residual network;

[0032] Step 0114) By integrating a spectral attention mechanism, the weights of different spectral bands are dynamically adjusted;

[0033] Step 0115) Output the reconstructed image quantum state data through quantum measurement.

[0034] As a further improvement to this embodiment of the invention, step 012) specifically includes:

[0035] Step 0121) Pre-train using a classic high-resolution hyperspectral dataset;

[0036] Step 0122) Construct a multi-scale dilated convolutional feature extraction network and obtain features at different scales by changing the dilation rate;

[0037] Step 0123) Configure a spectral feature compression layer to reduce the dimensionality of spectral information through 1×1 convolution;

[0038] Step 0124) Set the binary classification output, and output the probability that the image is a real image.

[0039] As a further improvement to this embodiment of the invention, step 20) specifically includes:

[0040] Step 201) Input the low-resolution hyperspectral data after quantum encoding in step 104) into the quantum generator constructed in step 011), and output the reconstructed image data;

[0041] Step 202) Input the reconstructed image data into the discriminator constructed using step 012) and output the probability that the image is a real image;

[0042] Step 203) Update network parameters by calculating the loss function;

[0043] Step 204) Determine if the image quality meets the output conditions. If it does, output a high-quality hyperspectral reconstructed image G*(I LR If it does not meet the requirements, return to steps 201) and 202) to regenerate and judge, and then update the parameters through step 203) until the output conditions are met.

[0044] As a further improvement to this embodiment of the invention, step 203) specifically includes:

[0045] Step 2031) Constructs the generator joint loss function shown in Equation (5) using adversarial loss, mean squared error loss, and spectral angle mapping loss:

[0046]

[0047] in, To combat the losses, For mean square error loss, Let λ1, λ2, and λ3 be the weighting coefficients for the spectral angle mapping loss. The spectral angle mapping loss function is shown in equation (6).

[0048]

[0049] In the formula, and δ represents the spectral vector of the i-th pixel in the real image and the reconstructed image, respectively. i Let be the spectral complexity factor of the i-th pixel, defined as the ratio of the variance of the pixel's spectral vector to the maximum variance;

[0050] Step 2032) The discriminator loss function is expressed as follows:

[0051]

[0052] In the formula, D(·) is the discriminator output, G(·) is the generator output, and I... HR For true high-resolution images, I LR Input images at low resolution;

[0053] Step 2033) Using the loss function obtained in steps 2031) and 2032), the discriminator parameters and generator parameters are updated. The discriminator parameter update is shown in Equation (8), and the generator parameter update is shown in Equation (9). By alternately optimizing the generator and discriminator parameters, a high-quality, high-resolution hyperspectral image is obtained.

[0054]

[0055] In the formula, η D The discriminant learning rate is t, and the iteration number is t.

[0056]

[0057] In the formula, η G The quantum circuit parameters are optimized using quantum gradient descent as the generator learning rate.

[0058] By reusing Equation (10), the learning rate is dynamically adjusted to produce higher quality reconstructed images;

[0059]

[0060] In the formula, η0 is the initial learning rate, and T is the decay constant.

[0061] As a further improvement to this embodiment of the invention, step 204) specifically includes:

[0062] Step 2041) First, use equation (11) to determine the convergence of the loss function and calculate the rate of change of the loss function.

[0063]

[0064] In the formula, This represents the rate of change of loss and measures the degree of training convergence. This represents the total generator loss in the t-th iteration. Let represent the total loss of the discriminator in the t-th iteration, where t represents the current iteration number;

[0065] Step 2042) First, use equation (12) to comprehensively evaluate the image quality and obtain the comprehensive image quality score Q;

[0066]

[0067] Step 2043) Perform output condition judgment. If the condition Q>Q is met, threshold and If the condition is met, output "Yes"; otherwise, output "No".

[0068] In the condition, Q t hreshold is the image quality threshold with a preset value of 0.90, and ∈ is the convergence threshold with a preset value of 10. -4 .

[0069] As a further improvement to this embodiment of the invention, step 30) specifically includes:

[0070] Step 301) Input the high-quality hyperspectral reconstructed image obtained in step 204) and perform image data preprocessing;

[0071] Step 302) Input the preprocessed image data into the quantum neural network for multi-level feature extraction. This quantum convolution operation represents:

[0072]

[0073] In the formula, U QCNN L is the transformation matrix of the quantum convolutional neural network. QNN For the number of layers in a quantum neural network, Let be the trainable parameters of the l-th layer quantum gate;

[0074] Step 303) Perform quantum classification on the extracted quantum features. The quantum variable classification circuit is as follows:

[0075]

[0076] In the formula, V(Θ) represents a variable component quantum circuit. Represents the set of trainable parameters;

[0077] Step 304) Perform POVM measurements on the quantum state to obtain the probability distribution of the target category:

[0078] p k =<ψ enhanced |M k |ψ enhanced > Equation (15)

[0079] In the formula, p k M represents the probability that a sample belongs to the k-th class. k Let ψ represent the POVM measurement operator for the k-th category. enhanced This refers to the quantum state after spectral feature enhancement processing;

[0080] Step 305) Output the final recognition result based on the measured probability to obtain the probability distribution of the target category.

[0081]

[0082] In the formula, The final predicted category label is represented by T, which represents the temperature parameter used to smooth out the effects of quantum noise, and K represents the number of different identification categories.

[0083] Step 306) Evaluate the recognition confidence using the recognition results. The confidence score is as follows:

[0084]

[0085] In the formula, the basic threshold for confidence is set to 0.85. If the result meets this threshold, the output meets the confidence level; otherwise, the output does not meet the confidence level.

[0086] Step 307) If the confidence level obtained in step 305) is lower than the threshold, return to step 302) to re-encode quantum encoding and feature extraction, and re-identify the target through steps 303) to 305) until the recognition confidence level meets the requirements;

[0087] Step 308) Output the repaired hyperspectral image, recognition results, and confidence level.

[0088] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0089] This invention provides a method for hyperspectral image reconstruction and target recognition based on quantum generative adversarial networks (GANs) and quantum neural networks. The method involves transforming a preprocessed hyperspectral image into its quantum state and inputting it into a GAN for feature extraction and image reconstruction. The generator parameters are optimized through adversarial training against a classical discriminator to obtain a high-quality reconstructed hyperspectral image. The reconstructed hyperspectral image is then input into the quantum neural network for multi-level quantum feature extraction and target recognition, outputting the recognition result and confidence level assessment. This invention effectively alleviates the gradient vanishing problem in quantum circuits through this architecture, while utilizing a spectral attention mechanism to improve the preservation of spectral features. It also enhances data adaptability through iterative optimization, significantly improving the accuracy and efficiency of hyperspectral image target recognition. Attached Figure Description

[0090] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0091] Figure 1 This is a flowchart of the hyperspectral image reconstruction and target recognition method based on quantum generative adversarial networks and quantum neural networks according to the present invention;

[0092] Figure 2 This is a schematic diagram of the structure of quantum generative adversarial network image reconstruction in the method of this embodiment of the invention;

[0093] Figure 3 This is a schematic diagram of the structure of quantum neural network image target recognition in the method of this embodiment of the invention; Detailed Implementation

[0094] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0095] This invention provides a method for reconstructing and identifying targets in hyperspectral images based on quantum generative adversarial networks and quantum neural networks, such as... Figure 1 As shown, it includes the following steps:

[0096] Step 10) Preprocess the hyperspectral image to be tested to obtain a standardized low-resolution hyperspectral image;

[0097] Step 20) Input the preprocessed image into a quantum generative adversarial network for image reconstruction to obtain a hyperspectral reconstructed image;

[0098] Step 30) Input the reconstructed hyperspectral image into the target recognition model based on quantum neural network, and output the repaired image and the recognition result.

[0099] The hyperspectral image reconstruction and recognition method of this invention first preprocesses the hyperspectral image to be tested, including calibration, cropping, Gaussian blurring, downsampling, and normalization, and converts it into a quantum state. Then, the hyperspectral data of the quantum state is input into a quantum generative adversarial network for reconstruction. The network parameters are optimized through adversarial training between the generator and the discriminator, and a high-quality hyperspectral reconstructed image is output. Finally, the reconstructed image is input into a quantum neural network for target recognition. Through multi-level quantum feature extraction and classification, the recognition result and confidence assessment are output.

[0100] The method of this invention effectively alleviates the gradient vanishing problem by combining quantum generative adversarial networks and quantum neural networks, improves the ability to preserve spectral features by utilizing spectral attention mechanisms, and enhances the adaptability of data through iterative optimization, thereby significantly improving the accuracy and efficiency of hyperspectral image target recognition.

[0101] Preferably, step 10) of the preprocessing in the method of this embodiment of the invention specifically includes:

[0102] Step 101) Crop the original input hyperspectral image to generate image blocks of uniform size;

[0103] Step 102) Perform Gaussian blur and downsampling on the cropped image to generate low-resolution training samples;

[0104] Step 103) Normalize the image data, mapping pixel values ​​to the range [-1, 1]. The normalization formula is as follows:

[0105]

[0106] In the formula, I represents the original image, and max(I) represents the maximum pixel value in the image. G Normalized image;

[0107] Step 104) The normalized hyperspectral image is converted into a quantum state through quantum encoding. The encoding function represents:

[0108]

[0109] In the formula, ψ in Input quantum state, I′ x，y，z The normalized hyperspectral pixel value at position (x, y, z), γ zLet |x>|y>|z> be the weighting factor for the z-th spectral band, where |x>|y>|z> represent the quantum basis of the spatial dimension and the spectral dimension, respectively. The denominator is a normalization factor that ensures that the modulus of the quantum state is 1.

[0110] Preferably, the quantum generative adversarial network in the method of the embodiments of the present invention, such as Figure 2 As shown, the following steps are used to construct it:

[0111] Step 011) Construct a generator for a quantum generative adversarial network, using a quantum residual network structure.

[0112] Specifically, it includes:

[0113] Step 0111) Input the quantum state information of the hyperspectral image after quantum encoding obtained in step 104);

[0114] Step 0112) Construct a multi-layer quantum circuit composed of quantum gates. The quantum convolution operation is represented as follows:

[0115]

[0116] In the formula, L gen The number of circuit layers in the quantum generator. Here are the training parameters for the i-th qubit in the l-th layer, where n is the total number of qubits, and CNOT is the training parameter for the i-th qubit. k,k+1 This is a controlled NOT gate operation with the control bit at k and the target bit at k+1;

[0117] Step 0113) Construct a quantum residual network and establish skip connections between convolutional layers. The connection mechanism is expressed by the following formula:

[0118] |ψ out >=αU(θ)|ψ in >+β|ψ in > Equation (4)

[0119] In the formula, α and β are the weight coefficients of the learnable quantum residual network;

[0120] Step 0114) By integrating a spectral attention mechanism, the weights of different spectral bands are dynamically adjusted;

[0121] Step 0115) Output the reconstructed image quantum state data through quantum measurement.

[0122] Step 012) Construct the discriminator of the quantum generative adversarial network, using a classic binary classification convolutional neural network structure.

[0123] Specifically, it includes:

[0124] Step 0121) Pre-train using a classic high-resolution hyperspectral dataset;

[0125] Step 0122) Construct a multi-scale dilated convolutional feature extraction network and obtain features at different scales by changing the dilation rate;

[0126] Step 0123) Configure a spectral feature compression layer to reduce the dimensionality of spectral information through 1×1 convolution;

[0127] Step 0124) Set the binary classification output, and output the probability that the image is a real image.

[0128] Step 20) The image reconstruction process specifically includes:

[0129] Step 201) Input the low-resolution hyperspectral data after quantum encoding in step 104) into the quantum generator constructed in step 011), and output the reconstructed image data;

[0130] Step 202) Input the reconstructed image data into the discriminator constructed using step 012) and output the probability that the image is a real image;

[0131] Step 203) Update network parameters by calculating the loss function;

[0132] Step 204) Determine if the image quality meets the output conditions. If it does, output a high-quality hyperspectral reconstructed image G*(I LR If it does not meet the requirements, return to steps 201) and 202) to regenerate and judge, and then update the parameters through step 203) until the output conditions are met.

[0133] Preferably, step 203) of updating network parameters specifically includes:

[0134] Step 2031) Constructs the generator joint loss function shown in Equation (5) using adversarial loss, mean squared error loss, and spectral angle mapping loss:

[0135]

[0136] in, To combat the losses, For mean square error loss, Let λ1, λ2, and λ3 be the weighting coefficients for the spectral angle mapping loss. The spectral angle mapping loss function is shown in equation (6).

[0137]

[0138] In the formula, and δ represents the spectral vector of the i-th pixel in the real image and the reconstructed image, respectively. i Let be the spectral complexity factor of the i-th pixel, defined as the ratio of the variance of the pixel's spectral vector to the maximum variance;

[0139] Step 2032) The discriminator loss function is expressed as follows:

[0140]

[0141] In the formula, D(·) is the discriminator output, G(·) is the generator output, and I... HR For true high-resolution images, I LR Input images at low resolution;

[0142] Step 2033) Use the loss function obtained in Step 2031) and Step 2032) to update the discriminator parameters and generator parameters. The discriminator parameter update is shown in Equation (8), and the generator parameter update is shown in Equation (9). By alternately optimizing the generator and discriminator parameters, a high-quality high-resolution hyperspectral image is obtained.

[0143]

[0144] In the formula, η D The discriminant learning rate is t, and the iteration number is t.

[0145]

[0146] In the formula, η G The quantum circuit parameters are optimized using quantum gradient descent as the generator learning rate.

[0147] By reusing Equation (10), the learning rate is dynamically adjusted to produce higher quality reconstructed images;

[0148]

[0149] In the formula, η0 is the initial learning rate, and T is the decay constant.

[0150] Step 204) Output condition judgment specifically includes:

[0151] Step 2041) First, use equation (11) to determine the convergence of the loss function and calculate the rate of change of the loss function.

[0152]

[0153] In the formula, This represents the rate of change of loss and measures the degree of training convergence. This represents the total generator loss in the t-th iteration. Let represent the total loss of the discriminator in the t-th iteration, where t represents the current iteration number;

[0154] Step 2042) First, use equation (12) to comprehensively evaluate the image quality and obtain the comprehensive image quality score Q;

[0155]

[0156] Step 2043) Perform output condition judgment. If the condition Q>Q is met, threshold and If the condition is met, output "Yes"; otherwise, output "No".

[0157] In the condition, Q threshold The preset value for the image quality threshold is 0.90, and the preset value for the convergence threshold is 10. -4 .

[0158] Preferably, in step 30) of the method of this embodiment, the target recognition model is a quantum neural network, such as... Figure 3 As shown, its identification process includes:

[0159] Step 301) Input the high-quality hyperspectral reconstructed image obtained in step 204) and perform image data preprocessing;

[0160] Step 302) Input the preprocessed image data into the quantum neural network for multi-level feature extraction. This quantum convolution operation represents:

[0161]

[0162] In the formula, U QCNN L is the transformation matrix of the quantum convolutional neural network. QNN For the number of layers in a quantum neural network, Let be the trainable parameters of the l-th layer quantum gate;

[0163] Step 303) Perform quantum classification on the extracted quantum features. The quantum variable classification circuit is as follows:

[0164]

[0165] In the formula, V(Θ) represents a variable component quantum circuit. Represents the set of trainable parameters;

[0166] Step 304) Perform POVM measurements on the quantum state to obtain the probability distribution of the target category:

[0167] p k =<ψ enhanced |M k |ψ enhanced Equation (15)

[0168] In the formula, pk M represents the probability that a sample belongs to the k-th class. k Let ψ represent the POVM measurement operator for the k-th category. enhanced This refers to the quantum state after spectral feature enhancement processing;

[0169] Step 305) Output the final recognition result based on the measured probability to obtain the probability distribution of the target category.

[0170]

[0171] In the formula, The final predicted category label is represented by T, which represents the temperature parameter used to smooth out the effects of quantum noise, and K represents the number of different identification categories.

[0172] Step 306) Evaluate the recognition confidence using the recognition results. The confidence score is as follows:

[0173]

[0174] In the formula, the basic threshold for confidence is set to 0.85. If the result meets this threshold, the output meets the confidence level; otherwise, the output does not meet the confidence level.

[0175] Step 307) If the confidence level obtained in step 305) is lower than the threshold, return to step 302) to re-encode quantum encoding and feature extraction, and re-identify the target through steps 303) to 305) until the recognition confidence level meets the requirements;

[0176] Step 308) Output the repaired hyperspectral image, recognition results, and confidence level.

[0177] The foregoing has shown and described 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 specific embodiments described above. The specific embodiments and descriptions in the specification are merely for further illustrating the principles of the present invention. Various changes and modifications can be made to the present invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed.

Claims

1. A method for reconstructing and recognizing targets from hyperspectral images based on quantum generative adversarial networks and quantum neural networks, characterized in that, Includes the following steps: Step 10) Preprocess the hyperspectral image to be tested to obtain a standardized low-resolution hyperspectral image; Step 20) Input the preprocessed image into a quantum generative adversarial network for image reconstruction to obtain a hyperspectral reconstructed image; Step 30) Input the reconstructed hyperspectral image into the target recognition model based on quantum neural network, and output the repaired image and the recognition result.

2. The method for reconstructing and recognizing targets from hyperspectral images based on quantum generative adversarial networks and quantum neural networks according to claim 1, characterized in that, Step 10) specifically includes: Step 101) Crop the original input hyperspectral image to generate image blocks of uniform size; Step 102) Perform Gaussian blur and downsampling on the cropped image to generate low-resolution training samples; Step 103) Normalize the image data, mapping pixel values ​​to the range [-1, 1]. The normalization formula is as follows: In the formula, I represents the original image, and max(I) represents the maximum pixel value in the image. G Normalized image; Step 104) The normalized hyperspectral image is converted into a quantum state through quantum encoding. The encoding function represents: In the formula, ψ in Input quantum state, I′ x，y,z The normalized hyperspectral pixel value at position (x, y, z), γ z Let |x>|y>|z> be the weighting factor for the z-th spectral band, where |x>|y>|z> represent the quantum basis of the spatial dimension and the spectral dimension, respectively. The denominator is a normalization factor that ensures that the modulus of the quantum state is 1.

3. The method for reconstructing and recognizing targets from hyperspectral images based on quantum generative adversarial networks and quantum neural networks according to claim 1, characterized in that, The quantum generative adversarial network specifically includes: Step 011) Construct a generator for a quantum generative adversarial network, using a quantum residual network structure; Step 012) Construct the discriminator of the quantum generative adversarial network, using a classic binary classification convolutional neural network structure.

4. The method for reconstructing and recognizing targets from hyperspectral images based on quantum generative adversarial networks and quantum neural networks according to claim 3, characterized in that, The generator in step 011) specifically includes: Step 0111) Input the quantum state information of the hyperspectral image after quantum encoding obtained in step 104); Step 0112) Construct a multi-layer quantum circuit composed of quantum gates. The quantum convolution operation is represented as follows: In the formula, L gen The number of circuit layers in the quantum generator. Here are the training parameters for the i-th qubit in the l-th layer, where n is the total number of qubits, and CNOT is the training parameter for the i-th qubit. k,k+1 This is a controlled NOT gate operation with the control bit at k and the target bit at k+1; Step 0113) Construct a quantum residual network and establish skip connections between convolutional layers. The connection mechanism is expressed by the following formula: |ψ out >=aU(θ)|ψ in >+β|ψ in > expression(4) In the formula, α and β are the weight coefficients of the learnable quantum residual network; Step 0114) By integrating a spectral attention mechanism, the weights of different spectral bands are dynamically adjusted; Step 0115) Output the reconstructed image quantum state data through quantum measurement.

5. The method for reconstructing and recognizing targets from hyperspectral images based on quantum generative adversarial networks and quantum neural networks according to claim 3, characterized in that, The discriminator in step 012) specifically includes: Step 0121) Pre-train using a classic high-resolution hyperspectral dataset; Step 0122) Construct a multi-scale dilated convolutional feature extraction network and obtain features at different scales by changing the dilation rate; Step 0123) Configure a spectral feature compression layer to reduce the dimensionality of spectral information through 1×1 convolution; Step 0124) Set the binary classification output, and output the probability that the image is a real image.

6. The method for reconstructing and recognizing targets from hyperspectral images based on quantum generative adversarial networks and quantum neural networks according to claim 1, characterized in that, Step 20) specifically includes: Step 201) Input the low-resolution hyperspectral data after quantum encoding in step 104) into the quantum generator constructed in step 011), and output the reconstructed image data; Step 202) Input the reconstructed image data into the discriminator constructed using step 012) and output the probability that the image is a real image; Step 203) Update network parameters by calculating the loss function; Step 204) Determine if the image quality meets the output conditions. If it does, output a high-quality hyperspectral reconstructed image G. * (I LR If it does not meet the requirements, return to steps 201) and 202) to regenerate and judge, and then update the parameters through step 203) until the output conditions are met.

7. The network optimization based on quantum generative adversarial networks according to claim 6, characterized in that, Step 203) specifically includes: Step 2031) Constructs the generator joint loss function shown in Equation (5) using adversarial loss, mean squared error loss, and spectral angle mapping loss: in, To combat the losses, For mean square error loss, Let λ1, λ2, and λ3 be the weighting coefficients for the spectral angle mapping loss. The spectral angle mapping loss function is shown in equation (6). In the formula, and δ represents the spectral vector of the i-th pixel in the real image and the reconstructed image, respectively. i Let be the spectral complexity factor of the i-th pixel, defined as the ratio of the variance of the pixel's spectral vector to the maximum variance; Step 2032) The discriminator loss function is expressed as follows: In the formula, D(·) is the discriminator output, G(·) is the generator output, and I... HR For true high-resolution images, I LR Input images at low resolution; Step 2033) Use the loss function obtained in Step 2031) and Step 2032) to update the discriminator parameters and generator parameters. The discriminator parameter update is shown in Equation (8), and the generator parameter update is shown in Equation (9). By alternately optimizing the generator and discriminator parameters, a high-quality high-resolution hyperspectral image is obtained. In the formula, η D The discriminant learning rate is t, and the iteration number is t. In the formula, η G The quantum circuit parameters are optimized using quantum gradient descent as the generator learning rate. By reusing Equation (10), the learning rate is dynamically adjusted to produce higher quality reconstructed images; In the formula, η0 is the initial learning rate, and T is the decay constant.

8. The network optimization based on quantum generative adversarial networks according to claim 6, characterized in that, Step 204) specifically includes: Step 2041) First, use equation (11) to determine the convergence of the loss function and calculate the rate of change of the loss function. In the formula, This represents the rate of change of loss and measures the degree of training convergence. This represents the total generator loss in the t-th iteration. Let represent the total loss of the discriminator in the t-th iteration, where t represents the current iteration number; Step 2042) First, use equation (12) to comprehensively evaluate the image quality and obtain the comprehensive image quality score Q; Step 2043) Perform output condition judgment. When the condition Q>Q is met... threshold and If the condition is met, output "Yes"; otherwise, output "No". In the condition, Q t hreshold is the image quality threshold with a preset value of 0.90, and ∈ is the convergence threshold with a preset value of 10. -4 .

9. The method according to claim 1, characterized in that, The target recognition model in step 30) is a quantum neural network, and its recognition process includes: Step 301) Input the high-quality hyperspectral reconstructed image obtained in step 204) and perform image data preprocessing; Step 302) Input the preprocessed image data into the quantum neural network for multi-level feature extraction. This quantum convolution operation represents: In the formula, U QCNN L is the transformation matrix of the quantum convolutional neural network. QNN For the number of layers in a quantum neural network, Let be the trainable parameters of the l-th layer quantum gate; Step 303) Perform quantum classification on the extracted quantum features. The quantum variable classification circuit is as follows: In the formula, V(Θ) represents a variable component quantum circuit. Represents the set of trainable parameters; Step 304) Perform POVM measurements on the quantum state to obtain the probability distribution of the target category: p k = (ψ enhanced |M k |ψ enhanced > Equation (15) In the formula, p k M represents the probability that a sample belongs to the k-th class. k Let ψ represent the POVM measurement operator for the k-th category. enhanced This refers to the quantum state after spectral feature enhancement processing; Step 305) Output the final recognition result based on the measured probability to obtain the probability distribution of the target category. In the formula, The final predicted category label is represented by T, which represents the temperature parameter used to smooth out the effects of quantum noise, and K represents the number of different identification categories. Step 306) Evaluate the recognition confidence using the recognition results. The confidence score is as follows: In the formula, the basic threshold for confidence is set to 0.

85. If the result meets this threshold, the output meets the confidence level; otherwise, the output does not meet the confidence level. Step 307) If the confidence level obtained in step 305) is lower than the threshold, return to step 302) to re-encode quantum and extract features, and re-identify the target through steps 303) and 305) until the recognition confidence level meets the requirements. Step 308) Output the repaired hyperspectral image, recognition results, and confidence level.

Images