Image classification method and system based on quantum network

Abstract

The application discloses a kind of based on quantum network's image classification method and system, including obtaining the quantum image classification network and training data set of training completion;Initial coprocessing quantum classifier is constructed and coprocessing quantum classifier is trained;Quantum denoising self-encoder is constructed and Gaussian noise is added in training data set to obtain noisy training data set;Noisy training data set is input to current quantum denoising self-encoder and the gap is calculated;Image output by quantum denoising self-encoder is input into coprocessing quantum classifier and loss is calculated;The quantum denoising self-encoder constructed is trained using gap and loss to obtain trained quantum denoising self-encoder;The anti-noise quantum image classification network of trained quantum denoising self-encoder and the quantum image classification network of training completion is connected in series to obtain, and image classification is carried out.The application not only realizes the image classification based on quantum network, but also has high reliability, good accuracy and good stability.

Description

Image classification method and system based on quantum networks Technical Field

[0001] This invention belongs to the field of quantum technology, specifically relating to an image classification method and system based on quantum networks. Background Technology

[0002] Image classification is a common application requirement; a stable and reliable image classification scheme can greatly ensure the reliability and accuracy of the classification results. Therefore, improving the stability and reliability of image classification schemes has always been one of the key research focuses for researchers.

[0003] Quantum neural networks have wide applications in quantum machine learning, and they also demonstrate superior performance in image classification tasks. However, like classical neural networks, quantum neural networks are susceptible to carefully designed micro-perturbations, which can lead to misclassification of examples carrying attacks, thus threatening the security and reliability of image classification models.

[0004] A denoising autoencoder is a framework for defending against adversarial examples in neural network image classifiers. MagNet is a representative model in classic machine learning. The advantage of this framework is that it neither modifies the protected classifier nor requires knowledge of the process of generating adversarial examples. When faced with minor perturbations, a denoising autoencoder can reduce the noise carried by the image, recovering the original image, thus enabling the image classification network to correctly classify the input image.

[0005] However, in the quantum realm, there is currently a lack of research on quantum denoising autoencoders to protect quantum classifiers against perturbations. This results in image classification schemes based on quantum networks having poor reliability and stability. Summary of the Invention

[0006] One of the objectives of this invention is to provide a quantum network-based image classification method that is highly reliable, accurate, and stable.

[0007] The second objective of this invention is to provide a system for implementing the aforementioned quantum network-based image classification method.

[0008] The image classification method based on quantum networks provided by this invention includes the following steps:

[0009] S1. Obtain the trained quantum image classification network and the corresponding training dataset;

[0010] S2. Construct an initial coprocessor quantum classifier and train the constructed initial coprocessor quantum classifier using the training dataset obtained in step S1 to obtain the coprocessor quantum classifier;

[0011] S3. Construct a quantum denoising autoencoder and add Gaussian noise to the training dataset to obtain a noisy training dataset;

[0012] S4. Input the noisy training dataset obtained in step S3 into the current quantum denoising autoencoder, and calculate the difference between the image output by the quantum denoising autoencoder and the corresponding original image without added Gaussian noise.

[0013] S5. Input the image output by the quantum denoising autoencoder in step S4 into the coprocessing quantum classifier obtained in step S2, and calculate the loss between the classification result output by the coprocessing quantum classifier and the label corresponding to the original image without Gaussian noise.

[0014] S6. Repeat steps S4 to S5, and use the gap obtained in step S4 and the loss obtained in step S5 to train the constructed quantum denoising autoencoder to obtain the trained quantum denoising autoencoder.

[0015] S7. Connect the trained quantum denoising autoencoder obtained in step S6 with the trained quantum image classification network to obtain a noise-resistant quantum image classification network, and use the obtained noise-resistant quantum image classification network for image classification.

[0016] Step S2, which involves constructing the initial coprocessor quantum classifier, specifically includes the following steps:

[0017] An initial coprocessor quantum classifier is constructed using an arbitrary quantum classifier network;

[0018] The arbitrary quantum classifier network includes quantum neural networks and quantum convolutional neural networks.

[0019] Step S3, which describes the construction of a quantum denoising autoencoder, specifically includes the following steps:

[0020] An existing quantum network is used as a quantum denoising autoencoder;

[0021] The existing quantum networks mentioned include Hardware-Efficient Ansatz (HEA), Strongly Entangling Layers (SEL), and Quantum Convolution Blocks (QCB).

[0022] The quantum denoising autoencoder constructed in step S3 includes the following steps:

[0023] The input image is embedded into network A1 using amplitude encoding; network A1 includes Hardware-Efficient Ansatz (HEA), Strongly Entangling Layers (SEL), and Quantum Convolution Block (QCB).

[0024] Measurement network A1 output quantum state The output quantum state is obtained. Phase information and the measurement probability of each ground state The phase information is rearranged according to the measurement probability and divided into four equal parts.

[0025] The information after being divided equally is embedded into intermediate networks B1 to B4 using amplitude encoding; among them, intermediate networks B1 to B4 include Hardware-Efficient Ansatz (HEA), Strongly Entangling Layers (SEL), and Quantum Convolution Block (QCB).

[0026] The output quantum states of intermediate networks B1~B4 are measured to obtain the phase information of the output quantum states and the measurement probabilities of each ground state; the phase information is rearranged according to the measurement probabilities and merged into one information path;

[0027] The obtained information is input into the third network C1; the third network C1 includes a Hardware-Efficient Ansatz (HEA), a Strongly Entangling Layer (SEL), and a Quantum Convolution Block (QCB).

[0028] The output quantum state of the third network C1 is measured to obtain the phase information of the output quantum state and the measurement probabilities of each ground state; the phase information is then rearranged according to the measurement probabilities and divided by... The classical image is obtained and used as the output of the quantum denoising autoencoder.

[0029] The quantum denoising autoencoder constructed in step S3 is physically implemented through the following steps:

[0030] The quantum denoising autoencoder includes an encoder and a decoder; both the encoder and the decoder include several PQC circuits that provide unitary variable networks; each PQC circuit includes several rotation layers and entanglement layers;

[0031] The rotation layer consists of a set of universal rotation gates on each quantum circuit to form a fully unitary U, denoted as ,in A single-bit trainable parameterized rotation gate that rotates about the Z-axis of a Bloch sphere. A single-bit trainable parameterized rotation gate for rotation about the Y-axis of a Bloch sphere. The first parameter to be trained. The second parameter to be trained. The third parameter to be trained; the combination of rotating doors can be arbitrarily changed;

[0032] The entanglement layer consists of CNOT gates.

[0033] Step S4 specifically includes the following steps:

[0034] The noisy training dataset obtained in step S3 Input to the current quantum denoising autoencoder;

[0035] Obtain the image output from the current quantum denoising autoencoder. ;

[0036] Computational quantum denoising autoencoder output image Compared with the original image without added Gaussian noise The gap between for ,in It is the square root of the sum of the squares of the elements of a vector, used to measure the Euclidean length of the vector.

[0037] Step S5 specifically includes the following steps:

[0038] The image output from the quantum denoising autoencoder in step S4 The input is fed into the coprocessor quantum classifier obtained in step S2;

[0039] Obtain the classification results output by the coprocessor quantum classifier;

[0040] Calculate the cross-entropy loss between the classification result output by the coprocessor quantum classifier and the label corresponding to the original image without added Gaussian noise. .

[0041] The training described in step S6 specifically includes the following steps:

[0042] The difference obtained using step S4 and the loss obtained in step S5 Construct the total loss function for ,in The first weight value is set. This is the set second weight value;

[0043] Using the constructed total loss function The quantum denoising autoencoder is trained.

[0044] This invention also provides a system for implementing the aforementioned quantum network-based image classification method, comprising a data acquisition module, a classifier training module, an encoder construction module, a gap calculation module, a loss calculation module, an encoder training module, and an image classification module; the data acquisition module, classifier training module, encoder construction module, gap calculation module, loss calculation module, encoder training module, and image classification module are connected in series; the data acquisition module is used to acquire the trained quantum image classification network and the corresponding training dataset, and upload the data information to the classifier training module; the classifier training module is used to construct an initial coprocessor quantum classifier based on the received data information, and train the constructed initial coprocessor quantum classifier using the acquired training dataset to obtain the coprocessor quantum classifier, and upload the data information to the encoder construction module; the encoder construction module is used to construct a quantum denoising autoencoder based on the received data information, and add Gaussian noise to the training dataset to obtain a noisy training dataset, and upload the data information to the gap calculation module; the gap calculation module is used to construct a quantum denoising autoencoder based on the received data information, and add Gaussian noise to the training dataset to obtain a noisy training dataset, and upload the data information to the gap calculation module; the gap calculation module is used to construct the ... The data information is processed as follows: First, the noisy training dataset is input into the current quantum denoising autoencoder. The difference between the output image of the quantum denoising autoencoder and the corresponding original image without Gaussian noise is calculated, and the data is uploaded to the loss calculation module. Second, the loss calculation module, based on the received data, inputs the output image of the quantum denoising autoencoder into the obtained coprocessor quantum classifier. It calculates the loss between the classification result output by the coprocessor quantum classifier and the label corresponding to the original image without Gaussian noise, and uploads the data to the encoder training module. Third, the encoder training module, based on the received data, trains the constructed quantum denoising autoencoder using the obtained difference and loss, obtaining the trained quantum denoising autoencoder, and uploads the data to the image classification module. Fourth, the image classification module, based on the received data, concatenates the trained quantum denoising autoencoder with the trained quantum image classification network to obtain a noise-resistant quantum image classification network, and uses the noise-resistant quantum image classification network for image classification.

[0045] The image classification method and system based on quantum networks provided by this invention not only achieves image classification based on quantum networks by constructing a coprocessor quantum classifier and a quantum denoising autoencoder and performing joint training, but also has higher reliability, better accuracy, and better stability. Attached Figure Description

[0046] Figure 1 is a schematic diagram of the method flow of the present invention.

[0047] Figure 2 is a schematic diagram of the processing flow of the quantum denoising autoencoder constructed in the method of the present invention.

[0048] Figure 3 is a schematic diagram of an embodiment of the unitary variable network in the quantum denoising autoencoder constructed in the method of the present invention.

[0049] Figure 4 is a schematic diagram of the functional modules of the system of the present invention. Detailed Implementation

[0050] Figure 1 shows a schematic flowchart of the method of the present invention: The image classification method based on quantum networks disclosed in this invention includes the following steps:

[0051] S1. Obtain the trained quantum image classification network and the corresponding training dataset;

[0052] S2. Construct an initial coprocessor quantum classifier and train the constructed initial coprocessor quantum classifier using the training dataset obtained in step S1 to obtain the coprocessor quantum classifier;

[0053] In practice, any quantum classifier network can be used to construct the initial coprocessor quantum classifier; for example, the commonly used quantum neural network QNN or quantum convolutional neural network QCNN can be used as the initial coprocessor quantum classifier.

[0054] In practice, the coprocessor quantum classifier is used to assist in the training of the quantum denoising autoencoder. The coprocessor quantum classifier is only used during the training process. After training is completed, the coprocessor quantum classifier is not needed when applying it to specific image classification tasks.

[0055] S3. Construct a quantum denoising autoencoder and add Gaussian noise to the training dataset to obtain a noisy training dataset;

[0056] In practice, the quantum denoising autoencoder can use existing quantum networks or be built by itself;

[0057] Among them, the quantum networks that can be selected include Hardware-Efficient Ansatz (HEA), Strongly Entangling Layers (SEL), and Quantum Convolution Block (QCB).

[0058] However, while current quantum network structures have a theoretical basis for processing single quantum states, they cannot achieve a perfect mapping relationship when processing quantum state datasets due to the limitation of unitary transformations. Therefore, their performance is not ideal when used as quantum state-to-quantum state processing networks, such as quantum denoising autoencoders. Thus, this invention also constructs its own quantum denoising autoencoder. The processing steps of this quantum denoising autoencoder include the following steps (as shown in Figure 2):

[0059] The input image is embedded into network A1 using amplitude encoding; network A1 includes Hardware-Efficient Ansatz (HEA), Strongly Entangling Layers (SEL), and Quantum Convolution Block (QCB).

[0060] Measurement network A1 output quantum state The output quantum state is obtained. Phase information and the measurement probability of each ground state The phase information is rearranged according to the measurement probability and divided into four equal parts.

[0061] The information after being divided equally is embedded into intermediate networks B1 to B4 using amplitude encoding; among them, intermediate networks B1 to B4 include Hardware-Efficient Ansatz (HEA), Strongly Entangling Layers (SEL), and Quantum Convolution Block (QCB).

[0062] The output quantum states of intermediate networks B1~B4 are measured to obtain the phase information of the output quantum states and the measurement probabilities of each ground state; the phase information is rearranged according to the measurement probabilities and merged into one information path;

[0063] The obtained information is input into the third network C1; the third network C1 includes a Hardware-Efficient Ansatz (HEA), a Strongly Entangling Layer (SEL), and a Quantum Convolution Block (QCB).

[0064] In practice, network A1, intermediate networks B1~B4 and third network C1 can be selected from efficient quantum networks, strongly entangled structures and quantum convolutional blocks for application. Currently, the more common approach is to select efficient quantum networks for all of them.

[0065] The output quantum state of the third network C1 is measured to obtain the phase information of the output quantum state and the measurement probabilities of each ground state; the phase information is then rearranged according to the measurement probabilities and divided by... The classical image is obtained and used as the output of the quantum denoising autoencoder;

[0066] The quantum denoising autoencoder constructed by this invention uses a multi-level quantum network and utilizes mid-course measurement and rearrangement methods to overcome the limitation of a single unitary change, while making full use of the performance of the quantum network.

[0067] The physical implementation of a self-built quantum denoising autoencoder includes the following steps:

[0068] The quantum denoising autoencoder includes an encoder and a decoder; both the encoder and the decoder include several PQC circuits that provide unitary variable networks (a structural example is shown in Figure 3); each PQC circuit includes several rotation layers and entanglement layers;

[0069] The rotation layer consists of a set of universal rotation gates on each quantum circuit to form a fully unitary U, denoted as ,in A single-bit trainable parameterized rotation gate that rotates about the Z-axis of a Bloch sphere. A single-bit trainable parameterized rotation gate for rotation about the Y-axis of a Bloch sphere. The first parameter to be trained. The second parameter to be trained. The third parameter to be trained; the combination of rotating doors can be arbitrarily changed;

[0070] The entanglement layer consists of CNOT gates;

[0071] S4. Input the noisy training dataset obtained in step S3 into the current quantum denoising autoencoder, and calculate the difference between the image output by the quantum denoising autoencoder and the corresponding original image without added Gaussian noise; specifically including the following steps:

[0072] The noisy training dataset obtained in step S3 Input to the current quantum denoising autoencoder;

[0073] Obtain the image output from the current quantum denoising autoencoder. ;

[0074] Computational quantum denoising autoencoder output image Compared with the original image without added Gaussian noise The gap between for ,in It is the square root of the sum of the squares of the elements of a vector, used to measure the Euclidean length of the vector.

[0075] The noisy data is not generated by any attack method, but rather Gaussian noise is used to summarize the effect of any additive attack. Since Gaussian noise tends to cluster at low perturbations, it also produces some samples that are very close to the original data. Therefore, the denoiser receives both noisy and near-original data simultaneously, allowing it to process both noisy and noise-free data concurrently. The quantum denoising autoencoder, through training, should minimize the gap... The goal is to reduce the noise level and restore the noisy data to its original state.

[0076] S5. Input the image output by the quantum denoising autoencoder in step S4 into the coprocessing quantum classifier obtained in step S2, and calculate the loss between the classification result output by the coprocessing quantum classifier and the label corresponding to the original image without added Gaussian noise; specifically including the following steps:

[0077] The image output from the quantum denoising autoencoder in step S4 The input is fed into the coprocessor quantum classifier obtained in step S2;

[0078] Obtain the classification results output by the coprocessor quantum classifier;

[0079] Calculate the cross-entropy loss between the classification result output by the coprocessor quantum classifier and the label corresponding to the original image without added Gaussian noise. ;

[0080] S6. Repeat steps S4 to S5, and use the gap obtained in step S4 and the loss obtained in step S5 to train the constructed quantum denoising autoencoder to obtain the trained quantum denoising autoencoder.

[0081] The training process includes:

[0082] The difference obtained using step S4 and the loss obtained in step S5 Construct the total loss function for ,in The first weight value is set. This is the set second weight value;

[0083] Using the constructed total loss function Training a quantum denoising autoencoder

[0084] S7. Connect the trained quantum denoising autoencoder obtained in step S6 with the trained quantum image classification network to obtain a noise-resistant quantum image classification network, and use the obtained noise-resistant quantum image classification network for image classification.

[0085] In practice, the trained quantum denoising autoencoder is used as the pre-network of the trained quantum image classification network. The input image to be classified is first processed by the quantum denoising autoencoder and then passed through the trained quantum image classification network to achieve image classification of the input image.

[0086] The effects of the method of the present invention will be illustrated below with reference to an embodiment:

[0087] The method of this invention is combined with existing unprotected schemes, Defense-GAN schemes, MagNet schemes, QNN with Noise Layer schemes, and Depolar schemes to process the HEA network simultaneously. This experiment is a three-class classification task experiment conducted on the MNIST handwritten digit dataset, with the input data carrying an FGSM attack.

[0088] Among them, the Defense-GAN scheme was proposed by Samangouei et al. in 2018 in "Defense-GAN: Protecting classifiers against adversarial attacks using generative models"; the MagNet scheme was proposed by Meng et al. in 2017 in "Magnet: a two-pronged defense against adversarial examples"; the QNN with Noise Layer scheme was proposed by Huang et al. in 2023 in "Enhancing adversarial robustness of quantum neural networks by adding noise layers"; and the Depolar scheme was proposed by Du et al. in 2021 in "Quantum noise protects quantum classifiers against adversaries".

[0089] The final comparison data is shown in Table 1:

[0090]

[0091] As can be seen from Table 1, the present invention demonstrates superiority in image processing: training is faster and higher defense accuracy is achieved under small-scale network structures (7-bit network is used in this embodiment, and the image size is 11*11), and the model accuracy is better than other existing solutions; this is sufficient to demonstrate the reliability of the present invention.

[0092] Figure 4 shows a schematic diagram of the functional modules of the system of the present invention: The system for implementing the image classification method based on quantum networks disclosed in this invention includes a data acquisition module, a classifier training module, an encoder construction module, a gap calculation module, a loss calculation module, an encoder training module, and an image classification module; the data acquisition module, classifier training module, encoder construction module, gap calculation module, loss calculation module, encoder training module, and image classification module are connected in series; the data acquisition module is used to acquire the trained quantum image classification network and the corresponding training dataset, and upload the data information to the classifier training module; the classifier training module is used to construct an initial coprocessor quantum classifier based on the received data information, and train the constructed initial coprocessor quantum classifier using the acquired training dataset to obtain the coprocessor quantum classifier, and upload the data information to the encoder construction module; the encoder construction module is used to construct a quantum denoising autoencoder based on the received data information, and add Gaussian noise to the training dataset to obtain a noisy training dataset, and upload the data information to the gap calculation module; gap calculation... The first module takes received data and inputs the noisy training dataset into the current quantum denoising autoencoder, calculates the difference between the output image of the quantum denoising autoencoder and the corresponding original image without added Gaussian noise, and uploads the data to the loss calculation module. The second module takes received data and inputs the output image of the quantum denoising autoencoder into the obtained coprocessing quantum classifier, calculates the loss between the classification result output by the coprocessing quantum classifier and the label corresponding to the original image without added Gaussian noise, and uploads the data to the encoder training module. The third module takes received data and trains the constructed quantum denoising autoencoder using the obtained difference and loss, obtaining the trained quantum denoising autoencoder, and uploads the data to the image classification module. The fourth module takes received data and concatenates the trained quantum denoising autoencoder with the trained quantum image classification network to obtain a noise-resistant quantum image classification network, and uses the noise-resistant quantum image classification network for image classification.

Claims

1. An image classification method based on quantum networks, characterized in that... The process includes the following steps: S1. Obtain the trained quantum image classification network and the corresponding training dataset; S2. Construct an initial coprocessor quantum classifier and train it using the training dataset obtained in step S1 to obtain the coprocessor quantum classifier; S3. Construct a quantum denoising autoencoder and add Gaussian noise to the training dataset to obtain a noisy training dataset; The constructed quantum denoising autoencoder process includes the following steps: embedding the input image into network A1 using amplitude encoding; wherein, network A1 includes an efficient quantum network, a strongly entangled structure, and quantum convolutional blocks; measuring the output quantum state of network A1. The output quantum state is obtained. Phase information and the measurement probability of each ground state The phase information is rearranged according to the measurement probability and divided into four equal parts. The divided information is then embedded into intermediate networks B1-B4 using amplitude encoding. Intermediate networks B1-B4 include an efficient quantum network, a strongly entangled structure, and quantum convolution blocks. The output quantum states of intermediate networks B1-B4 are measured to obtain the phase information of the output quantum states and the measurement probabilities of each ground state. The phase information is then rearranged according to the measurement probabilities and merged into a single information stream. The obtained information is input into a third network C1. The third network C1 includes an efficient quantum network, a strongly entangled structure, and quantum convolution blocks. The output quantum states of the third network C1 are measured to obtain the phase information of the output quantum states and the measurement probabilities of each ground state. The phase information is then rearranged according to the measurement probabilities and divided by... S3. Obtain the classical image and use it as the output of the quantum denoising autoencoder; S4. Input the noisy training dataset obtained in step S3 into the current quantum denoising autoencoder and calculate the difference between the image output by the quantum denoising autoencoder and the corresponding original image without Gaussian noise; S5. Input the image output by the quantum denoising autoencoder in step S4 into the coprocessing quantum classifier obtained in step S2 and calculate the loss between the classification result output by the coprocessing quantum classifier and the label corresponding to the original image without Gaussian noise; S6. Repeat steps S4 to S5, using the difference obtained in step S4 and the loss obtained in step S5 to train the constructed quantum denoising autoencoder, and obtain the trained quantum denoising autoencoder; S7. Connect the trained quantum denoising autoencoder obtained in step S6 with the trained quantum image classification network to obtain a noise-resistant quantum image classification network, and use the obtained noise-resistant quantum image classification network for image classification.

2. The image classification method based on quantum networks according to claim 1, characterized in that... The construction of the initial coprocessor quantum classifier in step S2 specifically includes the following steps: constructing the initial coprocessor quantum classifier using an arbitrary quantum classifier network; the arbitrary quantum classifier network includes quantum neural networks and quantum convolutional neural networks.

3. The image classification method based on quantum networks according to claim 2, characterized in that... The construction of the quantum denoising autoencoder described in step S3 specifically includes the following steps: using an existing quantum network as the quantum denoising autoencoder; the existing quantum network includes a hardware-efficient quantum network, a strongly entangled structure, and quantum convolutional blocks.

4. The image classification method based on quantum networks according to claim 3, characterized in that... The quantum denoising autoencoder constructed in step S3 is physically implemented as follows: The quantum denoising autoencoder includes an encoder and a decoder; both the encoder and decoder include several PQC circuits providing unitary networks; each PQC circuit includes several rotation layers and entanglement layers; the rotation layer consists of a set of universal rotation gates on each quantum circuit to form a fully unitary U, denoted as... ,in A single-bit trainable parameterized rotation gate that rotates about the Z-axis of a Bloch sphere. A single-bit trainable parameterized rotation gate for rotation about the Y-axis of a Bloch sphere. The first parameter to be trained. The second parameter to be trained. The third parameter to be trained; the combination of rotating gates can be arbitrarily changed; the entanglement layer is composed of CNOT gates.

5. The image classification method based on quantum networks according to claim 4, characterized in that... Step S4 specifically includes the following steps: processing the noisy training dataset obtained in step S3... Input the current quantum denoising autoencoder; obtain the image output by the current quantum denoising autoencoder. ； Computational quantum denoising autoencoder output image Compared with the original image without added Gaussian noise The gap between for ,in It is the square root of the sum of the squares of the elements of a vector, used to measure the Euclidean length of the vector.

6. The image classification method based on quantum networks according to claim 5, characterized in that... Step S5 specifically includes the following steps: processing the image output by the quantum denoising autoencoder in step S4... The input is fed into the coprocessor quantum classifier obtained in step S2; Obtain the classification result output by the coprocessor quantum classifier; calculate the cross-entropy loss between the classification result output by the coprocessor quantum classifier and the label corresponding to the original image without added Gaussian noise. 。 7. The image classification method based on quantum networks according to claim 6, characterized in that... The training described in step S6 specifically includes the following steps: using the gap obtained in step S4 and the loss obtained in step S5 Construct the total loss function for ,in The first weight value is set. The second weight value is set; the constructed total loss function is used. The quantum denoising autoencoder is trained.

8. A system for implementing the image classification method based on quantum networks as described in any one of claims 1 to 7, characterized in that... It includes a data acquisition module, a classifier training module, an encoder construction module, a gap calculation module, a loss calculation module, an encoder training module, and an image classification module; the data acquisition module, classifier training module, encoder construction module, gap calculation module, loss calculation module, encoder training module, and image classification module are connected in series; the data acquisition module is used to acquire the trained quantum image classification network and the corresponding training dataset, and upload the data information to the classifier training module; The classifier training module is used to construct an initial coprocessor quantum classifier based on the received data information, and to train the constructed initial coprocessor quantum classifier using the acquired training dataset to obtain the coprocessor quantum classifier, and then upload the data information to the encoder construction module; the encoder construction module is used to construct a quantum denoising autoencoder based on the received data information, and to add Gaussian noise to the training dataset to obtain a noisy training dataset, and then upload the data information to the gap calculation module. The gap calculation module is used to input the noisy training dataset into the current quantum denoising autoencoder based on the received data information, calculate the gap between the image output by the quantum denoising autoencoder and the corresponding original image without added Gaussian noise, and upload the data information to the loss calculation module. The loss calculation module is used to input the image output by the quantum denoising autoencoder into the obtained coprocessing quantum classifier based on the received data information, calculate the loss between the classification result output by the coprocessing quantum classifier and the label corresponding to the original image without added Gaussian noise, and upload the data information to the encoder training module; the encoder training module is used to train the constructed quantum denoising autoencoder based on the received data information, using the obtained gap and the obtained loss, to obtain the trained quantum denoising autoencoder, and upload the data information to the image classification module; The image classification module is used to connect the trained quantum denoising autoencoder with the trained quantum image classification network based on the received data information to obtain a noise-resistant quantum image classification network, and then use the obtained noise-resistant quantum image classification network to classify images.

Images