An image denoising method based on HQC-MCDCNN

By combining quantum and classical components, the HQC-MCDCNN model addresses the problem of excessive computational and storage resources in existing image denoising methods, achieving a more efficient improvement in image denoising performance, particularly in terms of significant improvements in RMSE, PSNR, SSIM, FSIM, SRE, and UIQ metrics.

CN118570090BActive Publication Date: 2026-11-17CHONGQING NORMAL UNIVERSITY
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202410690953.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-05-30
Publication Date
2026-11-17
Estimated Expiration
2044-05-30

AI Technical Summary

Technical Problem

Existing image denoising methods are computationally and storage-intensive and struggle to meet higher performance requirements, especially in terms of image reconstruction metrics such as RMSE, PSNR, SSIM, FSIM, SRE, and UIQ.

Method used

The HQC-MCDCNN model is adopted, which combines quantum and classical parts. The quantum part performs feature extraction through multi-scale variable quantum circuits, while the classical part performs supervised learning through a multi-path denoising convolutional neural network. The quantum part is used as a preprocessing step and does not participate in subsequent parameter updates, while the classical part performs image denoising.

Benefits of technology

It effectively reduces the burden of computing and storage resources while improving image denoising performance, especially in terms of RMSE, PSNR, SSIM, FSIM, SRE and UIQ metrics, with an average UIQ metric improvement of 56.5% compared to existing models.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118570090B_ABST
    Figure CN118570090B_ABST
Patent Text Reader

Abstract

This invention relates to the field of image denoising technology, specifically to an image denoising method based on HQC-MCDCNN, comprising the following: using an HQC-MCDCNN model for image denoising; the HQC-MCDCNN model includes: a quantum part for feature extraction via multi-scale variable quantum circuitry, and a classical part for supervised learning via a multi-path denoising convolutional neural network; the multi-scale variable quantum circuitry extracts features from the image, generating an n-channel quantum feature map; path one of the classical multi-path denoising convolutional neural network: inputting the quantum feature map, passing through four layers of EAM and two long hop connections, and then adjusting the shape of the reconstructed image through a two-dimensional transposed convolutional layer; path two is the same as path one, but does not adjust the shape of the reconstructed image; the reconstructed images obtained from the two paths are stitched together to obtain the reconstructed feature map, then convolved to output the final denoised image. This solution reduces the burden on storage and computational resources, has low complexity, and improves denoising performance.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of image denoising technology, specifically to an image denoising method based on HQC-MCDCNN. Background Technology

[0002] Image denoising is one of the key issues in image processing. With the deepening of research, denoising algorithms have been continuously improved. Many of these algorithms rely on the Bayesian perspective, starting with early L2-based regularization, robust statistics, the introduction of wavelets, and later the deployment of partial differential equations in imaging tasks.

[0003] In recent years, with the development of deep learning technology and its application in the field of image processing, it has brought about tremendous changes and improved the effect of image noise reduction, and models with superior performance have emerged continuously.

[0004] However, as the scale of deep learning grows, its development has become severely limited by available computing resources. How to alleviate the storage and computational burden of the large amount of computing resources required by traditional computer image processing has become a key research problem. While existing image denoising methods have excellent performance, they still cannot meet higher performance requirements, especially in terms of some image reconstruction metrics such as RMSE, PSNR, SSIM, FSIM, SRE, and UIQ.

[0005] Therefore, there is an urgent need for an image denoising method based on HQC-MCDCNN that can reduce the burden of storage and computing resources, has low complexity, and improve denoising performance to meet higher image denoising requirements. Summary of the Invention

[0006] The present invention aims to provide an image denoising method based on HQC-MCDCNN, which can reduce the burden of storage and computing resources, has low complexity, and improve denoising performance.

[0007] This invention provides the following basic solution: an image denoising method based on HQC-MCDCNN, comprising the following:

[0008] Image denoising is performed using the HQC-MCDCNN model;

[0009] The HQC-MCDCNN model includes a quantum part and a classical part. The quantum part uses multi-scale variable quantum circuits for feature extraction, while the classical part uses a multi-path noise reduction convolutional neural network for supervised learning. Image noise reduction is achieved through the combined action of the quantum and classical parts.

[0010] Multiscale variable quantum circuits are used as convolution kernels to extract features from the input image with raw noise, generating an n-channel quantum feature map for each image;

[0011] Classical multi-path denoising convolutional neural networks include: Path one and Path two;

[0012] Path one: Input quantum feature map, pass through four layers of EAM, and two long skip connections, and then through a two-dimensional transposed convolutional layer to adjust the shape of the reconstructed image;

[0013] Path two: The input image with original noise is processed through four layers of EAM and two long skip connections, and the output reconstructed image has the same shape as the input image.

[0014] The reconstructed image obtained from path one and the reconstructed image obtained from path two are concatenated to obtain the reconstructed feature map, and then convolution is performed to output the final denoised image.

[0015] Beneficial effects of the basic scheme: This scheme uses the HQC-MCDCNN model for image denoising. The HQC-MCDCNN model is a hybrid quantum-classical multi-path denoising convolutional neural network model that combines quantum methods with classical models. In the HQC-MCDCNN model, multi-scale variable quantum circuits are used as convolution kernels to extract features from the input noisy image, generating an n-channel quantum feature map for each image, which is then input into the constructed classical multi-path denoising convolutional neural network for image denoising.

[0016] Multi-scale variable quantum circuits are used as convolution kernels. In these quantum circuits, each quantum gate only performs operations on its adjacent qubits. Within the same convolution kernel, all quantum gates have the same adjustable parameters. This preserves the local connectivity and weight sharing characteristics of convolutional neural networks, playing a crucial role in extracting complex feature information throughout the network model. Considering that quantum computers are currently limited by the number of qubits and the depth of quantum circuits, they still face some challenges, especially in handling complex image processing tasks. Therefore, in this scheme, multi-scale variable quantum circuits are used as convolution kernels only as a preprocessing step and do not participate in subsequent parameter updates. This effectively extracts image features and significantly reduces the training time of HQC-MCDCNN.

[0017] The classic multi-path denoising convolutional neural network employs two paths: Path one: inputs a quantum feature map, passes through four layers of EAM and two long skip connections, and then through a two-dimensional transposed convolutional layer to adjust the shape of the reconstructed image; Path two: inputs the original noisy image, passes through four layers of EAM and two long skip connections, and outputs a reconstructed image with the same shape as the input image. The reconstructed images obtained from path one and path two are concatenated to obtain the reconstructed feature map, which is then convolved to output the final denoised image. Based on the MNIST and Fashion-MNIST datasets (input images) with varying degrees of noise (noise factor between 0.3 and 0.7), this approach improves upon existing image denoising models in terms of RMSE, PSNR, SSIM, FSIM, SRE, and UIQ. In particular, the average UIQ is 56.5% higher than the worst-performing model, indicating that the HQC-MCDCNN model possesses superior feature extraction capabilities and excellent denoising performance.

[0018] In summary, this solution reduces the burden on storage and computing resources, has low complexity, and improves noise reduction performance to meet higher image noise reduction requirements. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating an embodiment of an image denoising method based on HQC-MCDCNN according to the present invention;

[0020] Figure 2 This is a schematic diagram of the MNIST noise dataset in an embodiment of an image denoising method based on HQC-MCDCNN of the present invention;

[0021] Figure 3 This is a schematic diagram of the Fashion-MNIST noisy dataset in an embodiment of an image denoising method based on HQC-MCDCNN of the present invention;

[0022] Figure 4 This is a schematic diagram of the quantum feature extraction process in an embodiment of an image denoising method based on HQC-MCDCNN of the present invention;

[0023] Figure 5 This is a schematic diagram of the Bloch spherical representation of a single quantum bit in an embodiment of an image denoising method based on HQC-MCDCNN of the present invention;

[0024] Figure 6 This is a flowchart illustrating the angle encoding process in an embodiment of an image denoising method based on HQC-MCDCNN of the present invention.

[0025] Figure 7 This is a flowchart illustrating the variable quantum algorithm in an embodiment of the image denoising method based on HQC-MCDCNN of the present invention.

[0026] Figure 8 This is a schematic diagram of the network structure of a multi-path denoising convolutional neural network in an embodiment of an image denoising method based on HQC-MCDCNN of the present invention;

[0027] Figure 9 This is a schematic diagram of the k-fold cross-validation process in an embodiment of an image denoising method based on HQC-MCDCNN of the present invention;

[0028] Figure 10 This is a schematic diagram of the RMSE of DnCNN on Fashion-MNIST in an embodiment of the image denoising method based on HQC-MCDCNN of the present invention;

[0029] Figure 11 This is a schematic diagram of the SSIM measurement structure in an embodiment of an image denoising method based on HQC-MCDCNN of the present invention;

[0030] Figure 12 This is a schematic diagram showing some experimental results of different models on MNIST in an embodiment of the image denoising method based on HQC-MCDCNN of the present invention;

[0031] Figure 13 This is a schematic diagram showing some experimental results of different models on Fashion-MNIST in an embodiment of the image denoising method based on HQC-MCDCNN of the present invention;

[0032] Figure 14 This is a schematic diagram illustrating the noise reduction effects of different models on fashion-mnist-noise-0.7 in an embodiment of an image noise reduction method based on HQC-MCDCNN of the present invention.

[0033] Figure 15 This is a schematic diagram illustrating the noise reduction effects of different models on fashion-mnist-noise-0.7 in an embodiment of an image noise reduction method based on HQC-MCDCNN of the present invention.

[0034] Figure 16 This is a schematic diagram illustrating the training history of HQC-MCDCNN on different MNIST noisy datasets in an embodiment of the image denoising method based on HQC-MCDCNN of the present invention.

[0035] Figure 17This is a schematic diagram illustrating the training history of HQC-MCDCNN on different Fashion-MNIST noisy datasets in an embodiment of the image denoising method based on HQC-MCDCNN of the present invention. Detailed Implementation

[0036] The following detailed description illustrates the specific implementation method:

[0037] The basic implementation examples are as follows: Figure 1 As shown: An image denoising method based on HQC-MCDCNN, including the following:

[0038] Image denoising is performed using the HQC-MCDCNN model; HQC-MCDCNN: Hybrid Quantum-Classical Multipath Denoising Convolutional Neural Network;

[0039] The HQC-MCDCNN model includes a quantum part and a classical part. The quantum part uses multi-scale variable quantum circuits for feature extraction, while the classical part uses a multi-path noise reduction convolutional neural network for supervised learning. Image noise reduction is achieved through the combined action of the quantum and classical parts.

[0040] Multiscale variable quantum circuits are used as convolution kernels to extract features from the input image with raw noise, generating an n-channel quantum feature map for each image;

[0041] Classical multi-path denoising convolutional neural networks include: Path one and Path two;

[0042] Path one: Input quantum feature map, pass through four layers of EAM, and two long skip connections, and then through a two-dimensional transposed convolutional layer to adjust the shape of the reconstructed image;

[0043] Path two: The input image with original noise is processed through four layers of EAM and two long skip connections, and the output reconstructed image has the same shape as the input image.

[0044] The reconstructed image obtained from path one and the reconstructed image obtained from path two are concatenated to obtain the reconstructed feature map, and then convolution is performed to output the final denoised image.

[0045] The construction method of the above HQC-MCDCNN model includes:

[0046] S1. Obtain the image as input, divide the image into training set and validation set, and input them into the constructed HQC-MCDCNN model; the HQC-MCDCNN model includes: quantum part and classical part; the quantum part is a multi-scale variable quantum circuit for feature extraction, and the classical part is a multi-path denoising convolutional neural network for supervised learning. Image denoising is completed by the combined action of the quantum part and the classical part.

[0047] Specifically, in the actual construction and training process, due to the limitations of NISQ (Noisy Intermediate-Scale Quantum), the number of qubits that can be simulated is relatively small, and it is not yet able to handle large-sized images. Therefore, two commonly used datasets were selected for this experiment: MNIST (Modified National Institute of Standards and Technology database, a large-scale handwritten digit database) and Fashion-MNIST (a 10-category clothing classification dataset), both of which are image datasets. The original dataset does not contain any noise, so a noisy dataset was manually prepared: a random number array with the same shape as the input data was generated, and the random numbers followed a normal distribution with a mean of 0 and a standard deviation of 1; the random number array was multiplied by a noise factor and added to the original data, thus introducing random noise; the data after introducing random noise was restricted to the range of 0.0 to 1.0 to ensure that the image pixel values ​​are still within a reasonable range; the function returns an image array with added noise; the above data is the image data; as shown Figure 2 and Figure 3 The diagram shown is a schematic diagram after adding noise, where 0.3, 0.4, 0.5, 0.6, and 0.7 represent the added noise factors.

[0048] S2. Perform quantum multi-scale feature extraction on the image, including: setting multi-scale variable quantum circuits as convolution kernels in the HQC-MCDCNN model to extract features from the input image, and generating an n-channel quantum feature map for each image; in this embodiment, n=8;

[0049] Specifically, the multi-scale variable quantum circuit, as a quantum convolutional layer, consists of two quantum convolution kernels: a shallow variable quantum circuit and a deep variable quantum circuit, for quantum multi-scale feature extraction. Compared with classical convolutional layers, quantum convolutional layers can utilize the physical properties of quantum computing to achieve parallel computation, significantly improving the computational efficiency of traditional machine learning when processing massive amounts of data, and achieving efficient information processing. The quantum convolutional layer is composed of parameterized quantum filters, which are similar to the convolution kernels in classical convolutional layers. In the quantum convolutional layer, the parameterized quantum filters are used to extract features from all qubits in the local space of the data, generating feature maps. In the quantum circuit, each quantum gate only performs operations on its adjacent qubits. Within the same quantum convolutional layer, all quantum gates have the same adjustable parameters, which preserves the characteristics of local connectivity and weight sharing in convolutional neural networks, playing an important role in extracting complex feature information in the entire network model.

[0050] These advantages enable quantum convolutional layers to effectively extract image features, reduce network model complexity, and significantly improve computational efficiency. However, in the era of NISQ (Noisy Intermediate-Scale Quantum), quantum computers are limited by the number of qubits and the depth of quantum circuits. These limitations make it difficult to build unrestricted, sufficiently large-scale quantum computers under current technological conditions.

[0051] At the current technological level, quantum computers still face some challenges, especially in handling complex image processing tasks. Therefore, in this scheme, the quantum convolutional layer is only used as a preprocessing step and does not participate in subsequent parameter updates (i.e., step S3). This effectively extracts image features and significantly reduces the training time of HQC-MCDCNN.

[0052] like Figure 4 As shown, the quantum convolutional layer for quantum multi-scale feature extraction set in this embodiment specifically includes, in part, S2:

[0053] S201. Initialize the multi-scale variable quantum circuit by initializing all n / 2 qubits to |0>. In this embodiment, n / 2 = 4; where qubit refers to a quantum bit.

[0054] A bit is the smallest unit of information in classical computing. The state of a classical bit is either 0 or 1. The concept of a quantum bit is derived from the classical bit

[20] . Corresponding to the 0 and 1 of a classical bit, a quantum bit has two possible computational ground states |0> and |1>. The vector representation of the two computational ground states is as follows: and The difference between a qubit and a classical bit is that the state of a qubit is not limited to |0> or |1>, but can be a linear combination of |0> and |1>, which can be represented as:

[0055]

[0056] Where α and β are amplitudes, |α| 2 and |β| 2 Let |α| represent the probabilities of obtaining 0 and 1 when measuring a qubit, respectively. 2 +|β| 2 =1, the state after linear combination is usually a superposition state.

[0057] The Bloch sphere is used for single-qubit state visualization. The Bloch sphere representation of a single qubit is as follows: Figure 5 As shown, based on its normalization condition, it can be equivalently expressed as:

[0058]

[0059] Where θ∈[0,π],

[0060] Due to the common phase e of |0> and |1> iγ Since it has no observational significance, discarding the phase factor yields the following effective form:

[0061]

[0062] Based on θ and the state of a single quantum bit Two parameters can uniquely correspond to a point on the Bloch sphere. To describe a system composed of two or more qubits, the concept of the tensor product from linear algebra is introduced, with the notation [symbol missing]. Tensor products can combine two vector spaces to form a larger vector space. Taking a two-qubit system as an example, the two-qubit system composed of qubits |φ1>=α1|0>+β1|1> and |φ2>=α2|0>+β2|1> is as follows:

[0063]

[0064] The tensor product symbol is omitted and abbreviated as:

[0065] |φ1φ2>=α1α2|00>+α1β2|01>+α2β1|10>+α2β2|11>;

[0066] The amplitudes of the four ground states satisfy the condition: |α1α2| 2 +|α1β2| 2 +|α2β1| 2 +|α2β2|2 =1;

[0067] Similarly, the state of an n-qubit system is:

[0068]

[0069] In the above formula, {0,1} n Let |ω> represent the set of binary numbers with n bits, and |ω> represent the ground state corresponding to ω.

[0070] S202. Perform angle encoding. Using a sliding window based on pre-set padding and stride, sample the image sequentially, acquiring n / 2 pixels each time, and inputting them as angles into n / 2 R values. y ;

[0071] Where padding is used to fill the edges of the input during convolution to control the size and shape of the output feature map; Stride is the stride; R y It is a revolving door.

[0072] Angular encoding encodes classical input (i.e., classical data) and maps it to a high-dimensional Hilbert space. It's a fundamental encoding method for converting classical data into quantum states. The input data is treated as angles, and then a rotation gate is used to rotate the data around the x-axis or y-axis, mapping the classical data to quantum states. By mapping all classical data to quantum states, angular encoding is completed.

[0073]

[0074] Where |ψ> represents the encoded mapped quantum state, Represents the Kronecker product (tensor product), |0 n > indicates i represents the i-th number, usually starting from 1, and R represents the Rotation Gate;

[0075] Expanding the above equation, we obtain:

[0076]

[0077] Using a simple sequence [1,2,3,4] as input data, to facilitate observation and calculation, each input data point is multiplied by π for amplification, resulting in [π,2π,3π,4π]; the resulting quantum state |0000> has internal details as follows: Figure 6 As shown; obtain Where |ψ1> is the quantum state of the variable quantum circuit after passing through the angle coding layer.

[0078] S203. Use the U quantum gate to perform various linear transformations on the state vector to form a quantum circuit U(θ). Some or all of the quantum gates contain free parameters, where θ is the set of free parameters, i.e., the weights of the quantum convolution kernel; where the state vector is the quantum state.

[0079] Variable quantum circuits, also known as parameterized quantum circuits, consist of three parts:

[0080] A fixed initial state is usually a |0> state;

[0081] A quantum circuit U(θ) consists of a series of quantum gates, some or all of which contain free parameters, where θ is the set of free parameters;

[0082] Measurement operations used to obtain observable quantities Measurement operations can be performed at the end of one or more quantum circuits;

[0083] The expected value of the variable quantum circuit after measurement is:

[0084]

[0085] The expected value f(θ) can be post-processed, or f(θ) can be used directly to define the loss function of the task objective. The expected value of the variable quantum circuit can be changed by adjusting the free parameters θ = (θ1, θ2, ...) to minimize the loss function. The optimization process is usually iterative, obtaining the optimal set of parameters through continuous updates.

[0086] When using variable quantum circuits to process tasks, a free parameter θ and a set of non-free parameters x = (x1, x2, ...) can be input into the quantum circuit as parameters for a quantum gate, transforming classical information x into quantum information for processing. The process of embedding classical information x into the variable quantum circuit is called data encoding. After adding fixed parameters x, the expected value of the variable quantum circuit is:

[0087]

[0088] It is important to note that variable quantum circuits are not the same as variable quantum algorithms. In this scheme, the variable quantum algorithm = variable quantum circuits + optimizer, such as... Figure 7 As shown, the similarities and differences between the two are illustrated.

[0089] S204. Measure the output quantum state. Each qubit is measured, generating a vector and sending it to the corresponding quantum feature map. One qubit generates one quantum feature map, corresponding to one channel. After the sliding window has completed traversing the input image, the quantum feature map of the image is officially generated and serialized and saved to the disk for subsequent model training.

[0090] Specifically, in the design of quantum circuits and quantum computing, measurement is a crucial operation for acquiring information about a quantum system. Through measurement, quantum states are transformed into classical measurement results, thereby obtaining the desired computational results. When performing a measurement, a specific set of ground states needs to be selected; these ground states are typically fundamental measurement bases related to the properties or target of the quantum system being measured.

[0091] According to the measurement principle of quantum mechanics, when measuring a quantum system, the probability of the measurement result is described by a set of measurement operators; the measurement operators act on the state space of the quantum system and observe it. If the quantum system is in the state |ψ> before the measurement, the probability of the measurement result being m is calculated by the following equation (3.11):

[0092]

[0093] in Represents the measurement operator M m The conjugate transpose (Hermitian conjugate), Let represent the Hermitian conjugate product of the measurement operators, and let P(m) describe the probability that the measurement result is m when the quantum system is in state |ψ> before the measurement. After the measurement, the state of the system collapses into:

[0094]

[0095] The measurement operator must satisfy the following completeness equation:

[0096]

[0097] The completeness equations ensure that the sum of the products of the measurement operators equals the unit operator, and the sum of the probabilities of all possible measurement outcomes is 1.

[0098]

[0099] This holds true for all |ψ>, which is equivalent to the completeness equation. However, it is simpler to directly verify the completeness equation, so the completeness equation is used as a constraint.

[0100] Furthermore, the calculation of the size of the quantum feature map includes:

[0101]

[0102] Where H represents the height of the input image, W represents the width of the input image, and F... h and F w The height and width of the convolution kernel are represented by P, the padding size by S, and the stride by H. out and W out Indicates the height and width of the feature map;

[0103] The number of channels in the feature map depends on the number of qubits; one qubit maps to one feature map. In this embodiment, since two variable quantum circuits are used to extract features from the same image, each variable quantum circuit has 4 qubits, which is equivalent to one circuit having 8 qubits. Therefore, 8 quantum feature maps are generated, with a shape of (27, 27, 8). This means that after each input image passes through the quantum convolution layer, 8 different quantum feature maps are obtained, each feature map corresponding to one channel. This structure provides rich quantum feature information for subsequent classical models.

[0104] S2 uses quantum convolutional layers for preprocessing. There are significant differences between quantum convolutional layers based on quantum filters and those based on classical filters: The data processing methods differ. Quantum convolutional kernels process quantum data mapped from classical data; encoding classical data into quantum circuits is a necessary process, while classical convolutional kernels can directly process the original classical data. The convolution operations differ: classical convolutional filters perform an inner product operation with the input and then sum; while quantum convolutional kernels first encode the input classical data into quantum states and then perform matrix multiplication (quantum gates). Therefore, quantum convolutional filters can accelerate the convolution process due to their quantum parallelism. The nonlinear processing differs: classical convolutional layers typically require the introduction of additional nonlinear functions, while quantum convolutional layers do not require additional nonlinear functions because quantum measurements themselves are nonlinear. These differences give quantum convolutional layers unique advantages in processing image features, especially in utilizing quantum parallelism and nonlinearity.

[0105] S3. The multi-path denoising convolutional neural network constructed by training the extracted features includes: inputting the quantum feature map into the classical multi-path denoising convolutional neural network set in the HQC-MCDCNN model, and training the classical multi-path denoising convolutional neural network.

[0106] Classical multi-path denoising convolutional neural networks include: Path one and Path two;

[0107] Path one: Input quantum feature map, shape (27,27,8), passes through four layers of EAM (EkhakcenektAttektiok Nodule, enhanced attention module), such as... Figure 8 As shown, the image is reconstructed by two long skip connections and then a two-dimensional transposed convolutional layer to adjust the shape of the reconstructed image to (28,28,8).

[0108] Path two: Input the original noisy image with shape (28,28,1), pass through four layers of EAM and two longskip connections, and output the reconstructed image with the same shape as the input image, which is (28,28,1).

[0109] The reconstructed image (28,28,8) obtained from path one and the reconstructed image (28,28,1) obtained from path two are concatenated according to axis=3 to obtain the reconstructed feature map (28,28,9). Then, after passing through a 1×1 convolution, the final denoised image is output.

[0110] Furthermore, during training, a k-fold hierarchical sampling strategy is adopted to divide the training set into training and validation sets k times for cross-validation. After training, the weights of each fold are saved for averaging when evaluating the HQC-MCDCNN model.

[0111] Specifically, a k-fold hierarchical sampling strategy is adopted to divide the training set into k folds. The input shape of each fold training set is (K,H,W,C)=(5000,28,28,1), the shape of the validation set is (1000,28,28,1), and the shape of each image is (28,28,1); in this embodiment, k=6.

[0112] In the k-fold training process described above, a k-fold stratified sampling strategy is adopted, which divides the training set into training and validation sets k times for cross-validation. This is called k-fold cross-validation. To ensure the fairness of the comparison between models, k-fold cross-validation is used. Among them, Stratified K-Fold Cross Validation (SKFCV) is an extension of traditional cross-validation. Its main purpose is to ensure that each class is evenly distributed throughout the k-fold, effectively maintaining the relative class frequency in each training and validation fold when using stratified sampling instead of random sampling.

[0113] In this embodiment, the training set contains 60,000 images, while the test set contains 10,000 images. During model training, the training set X_train is evenly divided into 6 parts, each containing 10,000 images. One part is selected as the validation set, and the remaining five parts are used as the training set, forming a fold. During the entire cross-validation process, each fold is used as the validation set once, ensuring a comprehensive and accurate model evaluation. That is, for a dataset, dividing it into 6 folds is equivalent to having 6 datasets.

[0114] Throughout the training cycle, each fold underwent a fair 50 epochs of training, and the weights were independently saved after training. Therefore, for a dataset, six training runs were performed, each containing 50 epochs, for a total of six folds, reaching 300 epochs. After the entire training process, one weight was retained for each fold. These six independently saved weights were loaded back into the model (multi-path denoising convolutional neural network) and used to make independent predictions using the validation set X_test. Each prediction result was measured against y_test (including RMSE, PSNR, SSIM, FSIM, SRE, and UIQ). The average of these six metrics was calculated to obtain the average metrics: average RMSE, average PSNR, average SSIM, average FSIM, average SRE, and average UIQ.

[0115] The dataset partitioning and final metric process for 6-fold cross-validation, as follows: Figure 9 As shown, this ensures a comprehensive evaluation of model performance and provides accurate metrics.

[0116] K-fold cross-validation helps to improve the overall understanding of model performance, reduces randomness and chance in the evaluation process, and thus makes the model evaluation more reliable and robust; for example, with the input of the original Fashion-MNIST dataset without any noise, after 50 epochs of training and 6-fold cross-validation, such as... Figure 10 As shown, the training RMSE of DnCNN on Fashion-MNIST with 6-fold cross-validation shows significant differences in the situation faced by each fold. The validation RMSE in fold 3 and fold 4 exhibits violent oscillations. Although the RMSE of each fold is close to 0 after 50 epochs of training, their PSNR and SRE differ greatly, with the highest PSNR being 65388.254 dB higher than the lowest PSNR.

[0117] S4. The reconstruction performance of the HQC-MCDCNN model is evaluated using several image reconstruction metrics, including one or more of RMSE, PSNR, SSIM, FSIM, SRE, and UIQ.

[0118] RMSE:

[0119] RMSE (Root Meam Squared Error) is the average of the squared differences between the reconstructed image and the original image. The smaller the RMSE value, the smaller the difference between the reconstructed image and the original image, and the less distortion.

[0120] If the input image is X(i,j) and the predicted image is Y(i,j), the MSE between the two images is:

[0121]

[0122] Among them, RMSE x The mean square error of channel x is represented by X(i,j) and Y(i,j), which represent the pixel values ​​at the corresponding coordinates, respectively. H and W are the height and width of the image, respectively.

[0123] If calculating a three-channel color image, the MSE of each channel is summed:

[0124] RMSE RGB =RMSE R +RMSE G +RMSE B ;

[0125] R, G, and B represent the red, green, and blue channels, respectively.

[0126] PSNR:

[0127] RMSE depends on image intensity scaling. For 8-bit images (pixel values ​​in the range of 0-255), a mean square error of 100.0 results in severe image distortion to the naked eye. However, for 10-bit images (pixel values ​​between [0,1023]), the MSE is 100.0, which is almost imperceptible to the human eye. To address this issue, PSNR converts the MSE to a logarithmic scale.

[0128] PSNR (Peak Signal-to-Noise Ratio) is measured in dB. A higher PSNR value indicates lower distortion. The formula for calculating PSNR is:

[0129]

[0130] Where MSE stands for Mean Square Error (RMSE), and MAX... I This represents the maximum pixel value, or amplitude, of an image. For example, if each pixel in an image is represented by 8 bits, then MAX... I =(2 8 -1) = 255.

[0131] SSIM:

[0132] When comparing the error or similarity of two images, the most common methods are to use the mean squared absolute error (MAE) and the mean squared error (MSE).

[0133] However, these two loss functions cannot fully express the intuitive perception of images by the human visual system; two images may only differ in brightness, but their MSE losses can differ greatly; while for a blurry image and a sharp image, the MSE loss may be very small. Therefore, another method that better reflects the human visual system's judgment of the similarity between two images is needed, namely the Structural Similarity Index (SSIM).

[0134] Unlike MSE and PSNR, which measure absolute error, SSIM is a perceptual model, meaning it better aligns with human visual perception. If the two input images are x and y, their structural similarity is:

[0135] SSIM(x,y)=[l(x,y)] α [c(x,y)] β [s(x,y)] γ ,

[0136]

[0137] Where l(x,y) is used to compare the brightness of x and y, c(x,y) is used to compare the contrast of x and y, s(x,y) is used to compare the structure of x and y, and α>0, β>0, γ>0 are three relatively important parameters used to adjust l(x,y), c(x,y), and s(x,y); μ x and μ y σ x and σ y Let be the mean and standard deviation of x and y, respectively; σ xy Let x and y be the covariances. C1, C2 and C3 are constants used to maintain the stability of l(x,y), c(x,y) and s(x,y).

[0138] The higher the SSIM value, the greater the similarity between the two images.

[0139] The mean, standard deviation, and covariance are calculated as follows:

[0140]

[0141] like Figure 11 As shown, the calculation process of SSIM is as follows. In addition, SSIM has the following two characteristics: SSIM has symmetry, that is, SSIM(x,y)=SSIM(y,x); the value range of the SSIM exponent is 0 to 1. When SSIM=1, it means that the two images are completely identical, and when SSIM=0, it means that the two images are completely different.

[0142] FSIM:

[0143] FSIM (Feature Similarity Index) uses feature similarity for quality assessment;

[0144] Because the human visual system (HVS) perceives images based on some low-level features, while phase congruency (PC) can well characterize local structures.

[0145] Meanwhile, since PC is relatively invariant to changes in images, it is beneficial to extract stable features from images. However, sometimes changes in images can indeed affect the visual experience, so gradient magnitude (GM) is needed to compensate for this.

[0146] FSIM uses two complementary features: PC (phase consistency) and GM (gradient magnitude).

[0147] Phase consistency (PC) is a method for detecting image features. One of the important characteristics of phase consistency is that it is invariant to changes in light in the image, emphasizing the image's frequency domain features; phase consistency is invariant to contrast.

[0148] GM, the calculation of image gradients, is a common technique in digital image processing. Convolutional masks are used to represent gradient operators, and multiple convolutional masks are used to measure the gradient. If f(x) is an image, G... x and G y Let f(x) be its horizontal and vertical gradients, respectively. Then the magnitude of the gradient of f(x) is:

[0149] Image quality is evaluated by calculating the similarity between two images. Let the two images be... (output image) and Y (ground truth), with phase coherences of PC1 and PC2 respectively; from the two images The PC image extracted from Y and the GM images G1 and G2 extracted from the two images.

[0150] Based on PC1, PC2, G1, and G2, obtain the FSIM, specifically including:

[0151] calculate Similarity to Y:

[0152]

[0153] Where T1 is a positive constant, and S can be increased. PCThe stability of T1 is calculated based on the PC value. The above formula describes a measure for determining the similarity between two positive real numbers, ranging from 0 to 1;

[0154] Calculate the similarity between G1 and G2:

[0155]

[0156] T2 is a positive constant that depends on the dynamic range of the gradient magnitude. Both T1 and T2 are constants to facilitate the use of FSIM.

[0157] S PC and S G Combined together, calculate Y and Fusion similarity S L :

[0158] S L (x)=[S PC (x)] α ·[S G (x)] β ;

[0159] The parameters α and β are used to adjust the relative importance of PC and GM features. In this embodiment, they are set as follows: α = β = 1; FSIM is a normalized value between 0 and 1.

[0160] S is calculated according to the above formula. L This approach treats all pixel weights in the entire image as equal, therefore, it uses PC as the weighting factor. Furthermore, whether a location in the image is more important than other locations is positively correlated with its PC value. Thus, FSIM is:

[0161]

[0162] Among them, PC m (x) = max(PC1(x), PC2(x)), where Ω represents the entire image space domain.

[0163] SRE:

[0164] SRE (Signal to Reconstruction Error Ratio) is a metric used to measure the quality of signal reconstruction. SRE is more suitable for comparing errors between images with different brightness levels; the popular PSNR does not achieve the same effect because the peak intensity is constant. SRE is:

[0165]

[0166] Where, μ xIt is the average value of x, and the value of SRE is in decibels (dB).

[0167] UIQ:

[0168] UIQ (Universal Image Quality Index) is a metric used to evaluate the similarity between two images, and is used for quality assessment of images without a reference.

[0169] UIQ calculates the similarity between two images based on their correlation loss, luminance distortion, and contrast distortion. The value ranges from 0 to 1. A higher value indicates that the two images are more similar, a value of 1 indicates that they are completely identical, and a value of 0 indicates that the two images are completely different.

[0170] Specifically, if x and y are the reference image and the test image, then the UIQ is:

[0171]

[0172] The first term is the correlation coefficient between images x and y, the second term measures the similarity of the average brightness of the two images, and the third term measures the similarity of the contrast of the two images. When UIQ is close to 1, the two images are similar in structure and quality. Conversely, when it is close to 0, the two images are very different in structure and quality. UIQ is suitable for comparing grayscale images. For color images, UIQ needs to be calculated separately for each channel.

[0173] S5. Obtain and evaluate the HQC-MCDCNN model that meets the preset reconstruction performance requirements, and perform image denoising; the preset reconstruction performance requirements are set according to the actual needs, with the minimum threshold of the image reconstruction index.

[0174] The specific implementation process is as follows: Currently, image denoising is divided into three categories: filter-based methods, model-based methods, and learning-based methods. In recent years, deep network-based methods have become the mainstream method because they have achieved more promising denoising results than filter-based, model-based, and traditional learning-based methods. Therefore, in the comparative experiment, image denoising models based on the same learning method were selected for comparison. Further considering the fairness, this comparative experiment selected the most classic models, including: DnCNN, IRCNN, FSRCNN

[33] , SkipCAE, ResnetCAE, and InceptionCAE.

[0175] Before conducting the comparative experiments, to ensure fairness, the initialization conditions of all participating models were uniformly set to the values ​​in Table 1. The Glorot uniform initializer, also known as the Xavier uniform initializer, was used to initialize the convolutional kernel weights to ensure that the variance of each layer is as equal as possible, thereby better promoting the flow of information in the network. For image processing, MSE (MEAM Square Error) was chosen as the loss function; for the optimizer, the most commonly used Adam was adopted. Furthermore, based on experience, the learning rate was uniformly set to 0.001. For the datasets, different intensities of random noise were added to the original MNIST and Fashion-MNIST datasets, where noise-n represents the intensity factor of the random noise added to each image; for example, mnist-noise-0.3 represents the MNIST dataset with an added intensity factor of 0.3.

[0176] Table 1: Settings List

[0177]

[0178] The results obtained from the experiment are shown in Tables 2, 3, and 4 below. The average value of each index is taken:

[0179] Table 2: Experimental results of each model on the Fashion-MNIST dataset with noise factors of 0.3 and 0.4.

[0180]

[0181] Table 3: Experimental results of each model on the Fashion-MNIST dataset with noise factors of 0.5 and 0.6.

[0182]

[0183]

[0184] Table 4: Experimental results of each model on the Fashion-MNIST noise factor 0.7 dataset.

[0185]

[0186]

[0187] Based on the data in Tables 2, 3, and 4, HQC-MCDCNN performs best across all metrics, especially under conditions with the highest random noise, such as fashion-mnist-noise-0.7, where its advantage is even more pronounced. Compared to ResNetCAE, HQC-MCDCNN achieves a 37.5% reduction in AVG RMSE, a 27.9% improvement in AVG GPSNR, an 11.1% improvement in AVG FSIM, a 20.5% improvement in AVG GRE, and a 56.5% improvement in AVG UIQ.

[0188] The table above shows various metrics for the Fashion-MNIST noisy dataset. For the MNIST noisy dataset, such as... Figure 12 As shown, some experimental results of different models on MNIST are as follows: Figure 13 As shown, some experimental results of different models on Fashion-MNIST;

[0189] For the dataset with the most random noise in the experiment, fashion-mnist-noise-0.7, such as Figure 14 and Figure 15 As shown, the differences in noise reduction quality among various noise reduction models are illustrated, and an intuitive analysis of the noise reduction effect of each model is provided based on the images:

[0190] HQC-MCDCNN: The noise reduction effect is obvious. Although it cannot be completely identical to the ground truth, the contour features, smoothness, and number of noise points are extremely close to the ground truth, showing excellent noise reduction performance.

[0191] DnCNN: The noise reduction effect is good, but the outline is incomplete and there are a few noise points. It is slightly inferior to HQC-MCDCNN.

[0192] IRCNN and FSRCNN: Their noise reduction effect is similar to DnCNN, but their contour features are significantly missing and there is more noise. They are slightly inferior to HQC-MCDCNN.

[0193] SkipCAE and InceptionCAE: They have good noise reduction effects, close to HQC-MCDCNN, and perform well in contour features, smoothness and noise reduction, showing good performance in the model.

[0194] ResnetCAE: The noise reduction effect is the least ideal. It performs poorly in terms of contour features, smoothness, and noise, and is relatively worse than other models.

[0195] The training history of HQC-MCDCNN on MNIST and Fashion-MNIST datasets with different noise levels, such as Figure 16and Figure 17 As shown, in terms of RMSE volatility on the validation set, HQC-MCDCNN exhibits significant fluctuations on MNIST and Fashion-MNIST without added random noise. This phenomenon is reasonable because the original images input to the model have not undergone noise reduction processing, while the model shows a very stable trend in six-fold cross-training on images containing noise.

[0196] In terms of convergence speed, HQC-MCDCNN exhibits rapid convergence on MNIST and Fashion-MNIST with random noise; the model's loss decreases rapidly within the first 5 epochs, while it tends to level off after 20 epochs. This indicates that our model can adapt quickly to noisy data and rapidly reach a good training state.

[0197] In summary, this solution reduces the burden on storage and computing resources, has low complexity, and improves noise reduction performance to meet higher image noise reduction requirements.

[0198] The above descriptions are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. An image denoising method based on HQC-MCDCNN, characterized in that, Includes the following: Image denoising is performed using the HQC-MCDCNN model; The HQC-MCDCNN model includes a quantum part and a classical part. The quantum part uses multi-scale variable quantum circuits for feature extraction, while the classical part uses a multi-path noise reduction convolutional neural network for supervised learning. Image noise reduction is achieved through the combined action of the quantum and classical parts. Multiscale variable quantum circuits are used as convolution kernels to extract features from the input image with raw noise, generating an n-channel quantum feature map for each image; A classic multi-path denoising convolutional neural network includes: path one and path two; Path 1: Input quantum feature mapping, pass through four layers of EAM, and two long hop connections, and then through a two-dimensional transposed convolutional layer to adjust the shape of the reconstructed image; Path 2: The input image with original noise is processed through four layers of EAM and two long skip connections, and the output reconstructed image has the same shape as the input image. The reconstructed image obtained from path one and the reconstructed image obtained from path two are stitched together to obtain the reconstructed feature map, and then convolution is performed to output the final denoised image. The method for constructing the HQC-MCDCNN model includes: S1. Obtain the image as input, divide the image into training set and validation set, and input them into the constructed HQC-MCDCNN model; S2. Perform quantum multi-scale feature extraction on the image, including: setting up multi-scale variable quantum circuits as convolution kernels in the HQC-MCDCNN model to extract features from the input image and generating an n-channel quantum feature map for each image; S3. The multi-path denoising convolutional neural network constructed by training the extracted features includes: inputting the quantum feature map into the HQC-MCDCNN model to set up the classic multi-path denoising convolutional neural network, and training the classic multi-path denoising convolutional neural network. S2 includes: S201. Initialize the multi-scale variable quantum circuit, initializing all n / 2 qubits to... ; where qubit refers to a quantum bit; S202. Perform angle encoding. Using a sliding window based on pre-set padding and stride, sample the image sequentially, acquiring n / 2 pixels each time, and inputting them as the angle into the n / 2 pixels. ; Where padding is used to fill the edges of the input during convolution to control the size and shape of the output feature map; stride is the stride. It is a revolving door; S203. Using U-gates to perform various linear transformations on the state vector to construct quantum circuits. Some or all quantum gates contain free parameters. It is a set of free parameters; S204. Measure the output quantum state. Each qubit is measured, generating a vector and sending it into the corresponding quantum feature map; one qubit generates one quantum feature map, corresponding to one channel.

2. The image denoising method based on HQC-MCDCNN according to claim 1, characterized in that, The step S1 of obtaining the input image includes: obtaining image datasets from MNIST and Fashion-MNIST as input images; and preparing a noise dataset includes: generating a random number array with the same shape as the input image, wherein the random numbers follow a normal distribution with a mean of 0 and a standard deviation of 1; multiplying the random number array by a noise factor and adding it to the input image.

3. The image denoising method based on HQC-MCDCNN according to claim 1, characterized in that, During the S3 training process, a k-fold hierarchical sampling strategy is adopted to divide the training set into a training set and a validation set k times for cross-validation. After training, the weights of each fold are saved for averaging when evaluating the HQC-MCDCNN model.

4. The image denoising method based on HQC-MCDCNN according to claim 1, characterized in that, The method for constructing the HQC-MCDCNN model further includes: S4, evaluating the reconstruction performance of the HQC-MCDCNN model using several image reconstruction metrics, wherein the image reconstruction metrics include one or more of RMSE, PSNR, SSIM, FSIM, SRE, and UIQ.

5. The image denoising method based on HQC-MCDCNN according to claim 4, characterized in that, The RMSE mentioned: RMSE is the average of the squares of the pixel differences between the reconstructed image and the original image. The smaller the RMSE value, the smaller the difference between the reconstructed image and the original image, and the less distortion. If the input image is The predicted image is The MSE between the two images is: ; in, Indicates channel Mean square error and These represent the pixel values ​​at the corresponding coordinates. and These are the height and width of the image, respectively; If calculating a three-channel color image, the MSE of each channel is summed: ; R, G, and B represent the red, green, and blue channels, respectively. The PSNR: PSNR, a higher value indicates less distortion. PSNR is: ; Where MSE stands for Mean Squared Error (RMSE). This represents the maximum pixel value of an image; The SSIM: If the two input images are x and y respectively, then their structural similarity is: in, For comparison and brightness, For comparison and Contrast For comparison and The structure, It is used for adjustment Three relatively important parameters; and , and They are respectively and The mean and standard deviation; for and covariance, , and They are all used for maintenance A constant for stability; The higher the SSIM value, the higher the similarity between the two images; The mean, standard deviation, and covariance are calculated as follows: The FSIM: FSIM uses two complementary features: PC and GM. like It is an image. and Let these be its horizontal and vertical gradients, respectively. The gradient magnitude is: ; Image quality is evaluated by calculating the similarity between two images. Let the two images be... and The phase consistency is respectively and From two images and PC image extracted from the two images and GM image extracted from the two images. and ; according to , , and To obtain an FSIM, including: calculate and Similarity: ; in, It is a positive number, to increase stability, Calculated based on PC value; the above formula describes a measure for determining the similarity between two positive real numbers, ranging from 0 to 1; calculate and Similarity: ; in It is a positive constant, depending on the dynamic range of the gradient magnitude; and All are constants; Will and Integrated together, computation and Fusion similarity : ; Where parameters and Used to adjust the relative importance of PC and GM features, set to: FSIM is a normalized value between 0 and 1. FSIM is: ; in, , Represents the entire image spatial domain; The SRE: SRE is a metric used to measure the quality of signal reconstruction. SRE is particularly suitable for comparing errors between images of different brightness levels. SRE is: ; in, It is the average value of x, and the value of SRE is in decibels; The UIQ mentioned: UIQ is a metric used to evaluate the similarity between two images; UIQ calculates the similarity between two images based on their correlation loss, brightness distortion, and contrast distortion. The value ranges from 0 to 1. A higher value indicates that the two images are more similar, a value of 1 indicates that they are completely identical, and a value of 0 indicates that the two images are completely different. If x and y are the reference image and the test image, then the UIQ is: ; The first term is the correlation coefficient between images x and y, the second term measures the similarity of the average brightness of the two images, and the third term measures the similarity of the contrast of the two images. When UIQ is close to 1, the two images are similar in structure and quality. Conversely, when it is close to 0, the two images are very different in structure and quality.

6. The image denoising method based on HQC-MCDCNN according to claim 5, characterized in that, The method for constructing the HQC-MCDCNN model further includes: S5, obtaining an HQC-MCDCNN model that meets the preset reconstruction performance requirements and performing image denoising.

Citation Information

Patent Citations

  • Knee osteoarthritis grading method based on quantum-to-classical transfer learning

    CN116664931A

  • Breast cancer detection method based on variable component sub-tensor network transfer learning

    CN117333754A