Quantum feature fusion method for improving quantum convolutional neural network

Abstract

The invention relates to a quantum feature fusion method for improving a quantum convolutional neural network. The method comprises the following steps: firstly, performing data enhancement on input data and expanding training samples; then, inputting the enhanced data into a classic convolutional layer, and extracting local features; thirdly, performing quantum feature mapping on features extracted by the classical convolutional layer by using a quantum convolutional layer, and mapping the features to a high-dimensional Hilbert space; in a quantum feature fusion module, quantum features extracted by different quantum convolutional layers are fused through quantum circuit design, and efficient fusion of the features is realized. And finally, inputting the fused quantum features into a full connection layer to realize a classification or regression task. According to the method, the advantages of quantum calculation are fully utilized, and the feature extraction and fusion capability of the convolutional neural network is effectively improved through quantum feature fusion. The effectiveness of the method is verified through specific experiments.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of quantum computing and artificial intelligence, and specifically relates to a method for improving the classification performance of a quantum convolutional neural network by using quantum feature fusion technology. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, convolutional neural networks (CNN) have become an important tool in computer vision, speech recognition, natural language processing, and other fields. Traditional CNNs are mainly based on classical computer architecture, extracting features through convolution, pooling, and other operations, and using fully connected layers for classification or prediction. However, for applications such as intelligent video surveillance, vehicle autonomous driving, robot environmental perception, and visual human-computer interaction, as data size and model complexity increase, traditional CNNs face challenges such as low computational efficiency and insufficient generalization ability. In addition, the performance of classical computers is gradually approaching the physical limit, making it difficult to meet the growing computing demands. Therefore, exploring new computing paradigms and algorithm architectures has become an important research direction in the field of artificial intelligence.

[0003] Quantum computing, as a new computing mode, has the advantages of strong parallelism and high computational efficiency, providing a new approach to solving the limitations of classical CNNs. In recent years, research on quantum machine learning has made significant progress, and quantum neural networks (QNN) as a representative model have shown broad application prospects by combining quantum computing and classical neural networks. However, existing QNN methods mostly use simple quantum circuit structures, lacking sufficient utilization of quantum features, resulting in limited model expressiveness and inaccurate image target classification. In addition, existing methods usually design quantum circuits and classical neural networks separately, lacking effective fusion mechanisms, making it difficult to leverage the potential advantages of quantum computing. Therefore, for computer vision applications, how to design advanced quantum convolutional neural network architectures, fully utilize quantum features, achieve quantum-classical fusion, and improve model performance to enhance target recognition and classification accuracy in computer vision is a key problem that needs to be solved. SUMMARY

[0004] The present application provides an advanced quantum convolutional neural network method based on quantum feature fusion, which utilizes the ability of quantum computing to surpass classical computing, extracts and fuses quantum features, and constructs an efficient quantum convolutional neural network model. This method can significantly improve model performance, accelerate training and inference processes, and reduce computational resource consumption.

[0005] The technical solution adopted by the present application to achieve the above-mentioned purposes is as follows: a method for improving quantum feature fusion of a quantum convolutional neural network, comprising the following steps:

[0006] (1) performing enhancement processing on the input training data set to generate a plurality of sample data after data volume expansion;

[0007] (2) inputting the preprocessed sample data into a convolutional neural network to extract local features of the image and generate a feature map;

[0008] (3) converting the feature map into a quantum state and inputting it into a quantum convolutional layer to perform convolution operation on the input quantum state through a quantum circuit to generate a quantum feature map;

[0009] (4) inputting the quantum feature map extracted by the quantum convolutional layer into a quantum feature fusion module to fuse quantum features of different levels and scales to generate a quantum feature representation;

[0010] (5) inputting the quantum feature representation output by the quantum feature fusion module into a fully connected layer through a measurement operation to map the quantum feature to an output space to obtain a prediction result.

[0011] The enhancement processing on the input training data set to generate a plurality of sample data after data volume expansion includes the following steps:

[0012] randomly rotating the input image data, with the rotation angle randomly selected between 0 and 360 degrees;

[0013] performing Gaussian blur processing on the rotated image data;

[0014] occluding at a random position of the image data, with the occluded area accounting for 5% to 15% of the entire image data, and the pixel value of the occlusion being randomly generated.

[0015] The conversion of the feature map into a quantum state and the inputting into a quantum convolutional layer to perform convolution operation on the input quantum state through a quantum circuit to generate a quantum feature map includes the following steps:

[0016] 1) mapping the feature map data to a quantum state space through an encoding module E; using a rotation gate R y (θ) and a Hadamard gate H to encode the input data into a quantum state;

[0017] 2) using a variational quantum circuit module U to transform and extract features from the encoded quantum state;

[0018] 3) measuring the quantum state output by the variational quantum circuit module U through a measurement module M to convert and output the quantum state as a quantum feature map.

[0019] The inputting of the quantum feature map extracted by the quantum convolutional layer into a quantum feature fusion module to fuse quantum features of different levels and scales to generate a quantum feature representation includes the following steps:

[0020] The plurality of quantum features of the quantum feature map is mapped to a Hilbert space, the probability amplitudes of each quantum feature are linearly combined by using the quantum superposition principle, a superposition state is constructed, and fusion of the quantum features in the Hilbert space is realized;

[0021] The fusion coefficient formula is represented as:

[0022] f′1(w1)=f1(w1)+a1f1(w1),

[0023] fv2(w2)=f2(w2)+a2f2(w2),

[0024] f′3(w3)=f3(w3)+a3f3(w3),

[0025] f′4(w4)=f4(w4)+a4f4(w4),

[0026] f′(W)=(a1f1(w1)+a2f2(w2)+a3f3(w3)+a4f4(w4)) / 4,

[0027]

[0028] wherein f i (w i ) is the i-th layer quantum feature map, w i is the weight used for generating the quantum feature map f i (w i ) by the quantum convolution layer, a i is the weight coefficient of f i (w i ), f′ i (w i ) is the i-th layer feature map after weighted fusion according to a i , f′(W) is a new feature map obtained by adding and fusing f′ i (w i ), and W={w1,w2,w3,w4}; is a new feature map obtained by channel superposition fusion of f′(W) and f′ i (w i ), that is, a quantum feature representation.

[0029] The quantum feature representation output by the quantum feature fusion module is input into a fully connected layer through a measurement operation to map the quantum feature to an output space and obtain a prediction result, which is realized by the following formula:

[0030] C F =f(Q⊙W2+b2)=σ(Q⊙W2+b2)

[0031] C in the above formula F is a fully connected layer, Q is a quantum feature representation output by a quantum feature fusion module; W2 is a weight matrix, b2 is a bias matrix; sigma is a ReLU activation function; and is a convolution operation.

[0032] A system for improving quantum feature fusion of a quantum convolutional neural network, comprising:

[0033] A data augmentation module is configured to perform augmentation processing on an input training data set to generate a plurality of sample data after data volume expansion.

[0034] A convolutional neural network module is configured to input the preprocessed sample data into a convolutional neural network to extract local features of the image and generate a feature map.

[0035] A quantum convolutional layer is configured to convert the feature map into a quantum state and input the quantum state into the quantum convolutional layer to perform convolution operation on the input quantum state through a quantum circuit to generate a quantum feature map.

[0036] A quantum feature fusion module is configured to input the quantum feature map extracted by the quantum convolutional layer into the quantum feature fusion module to fuse quantum features of different levels and scales to generate a quantum feature representation.

[0037] A fully connected layer is configured to input the quantum feature representation output by the quantum feature fusion module into the fully connected layer through a measurement operation to map the quantum feature to an output space to obtain a prediction result.

[0038] An apparatus for improving quantum feature fusion of a quantum convolutional neural network, comprising a memory and a processor; the memory is configured to store a computer program; the processor is configured to implement the method for improving quantum feature fusion of a quantum convolutional neural network when the computer program is executed.

[0039] A computer readable storage medium, the storage medium stores a computer program, when the computer program is executed by a processor, the method for improving quantum feature fusion of a quantum convolutional neural network is realized.

[0040] The present application has the following advantages and benefits:

[0041] 1. The present application proposes an innovative quantum convolutional neural network architecture, which introduces quantum feature extraction and fusion modules to fully utilize the parallelism and high-dimensional characteristics of quantum computing, significantly improving the feature expression ability and generalization performance of the model. The quantum feature extraction module adopts a parameterized quantum circuit, which can generate high-dimensional quantum states to capture the nonlinear and high-order statistical characteristics of the data. The quantum feature fusion module realizes information interaction and fusion of quantum features between different quantum bits through entanglement operation, enhances the semantic representation ability of the features, and is beneficial to the target classification of images.

[0042] 2. The application adopts a novel quantum-classical hybrid training paradigm, which realizes the end-to-end joint training of quantum circuits and classical neural networks through the gradient backpropagation algorithm and parameter optimization of the parameterized quantum circuit. This fusion mechanism allows quantum computing and classical computing to promote and optimize each other, fully utilizes the advantages of quantum computing, overcomes the difficulties of pure quantum circuit training, improves the training efficiency and performance of the model, and further improves the recognition and classification accuracy of targets in computer vision. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 The method flowchart of the application is shown in the figure;

[0044] Figure 2 The training error graph of the classical convolutional neural network (CNN), quantum convolutional neural network (QNN), and advanced quantum convolutional neural network (QNN-F) with quantum feature fusion module on the MNIST dataset is shown in the figure;

[0045] Figure 3 The test accuracy graph of the classical convolutional neural network (CNN), quantum convolutional neural network (QNN), and advanced quantum convolutional neural network (QNN-F) with quantum feature fusion module on the MNIST dataset is shown in the figure;

[0046] Figure 4 The training error graph of the classical convolutional neural network (CNN), quantum convolutional neural network (QNN), and advanced quantum convolutional neural network (QNN-F) with quantum feature fusion module on the Cifar-10 dataset is shown in the figure;

[0047] Figure 5 The test accuracy graph of the classical convolutional neural network (CNN), quantum convolutional neural network (QNN), and advanced quantum convolutional neural network (QNN-F) with quantum feature fusion module on the Cifar-10 dataset is shown in the figure. DETAILED DESCRIPTION

[0048] The application will be further described in detail below in combination with the drawings and examples.

[0049] The application provides a method for improving quantum feature fusion of a quantum convolutional neural network, which is applied to intelligent video monitoring, vehicle automatic driving, robot environment perception, visual human-computer interaction and the like. First, data augmentation is performed on original data to expand the diversity of training samples. Then, a classical convolutional layer extracts local features, and a quantum convolutional layer extracts global features by using superposition of quantum bits, so that classical-quantum feature fusion is realized. Finally, a fully connected layer integrates feature information and outputs a prediction result. The method fully utilizes the advantages of classical convolution and quantum convolution, significantly improves the feature expression capability and generalization performance of the model through multi-level feature extraction and fusion, and further improves the recognition and classification precision of a target in computer vision.

[0050] As shown in Figure 1 A method for improving quantum feature fusion of a quantum convolutional neural network comprises the following steps:

[0051] Step 1: Data augmentation. For an input training data set, data augmentation technology is adopted to generate multiple augmented samples by performing rotation, translation, scaling, flipping and the like on original images, expand the training data volume, and improve the generalization capability and robustness of the model. The original image is a real-time image collected by a video sensor or an image stored in a database.

[0052] Step 2: The augmented data is input into a classical convolutional neural network, and the local features of the image are extracted through a convolutional layer. Each convolutional layer is composed of multiple convolutional kernels, and a feature map is generated through convolution operation and a nonlinear activation function.

[0053] Step 3: The feature map extracted by the classical convolutional layer is converted into a quantum state and input into a quantum convolutional layer. The quantum convolutional layer is composed of multiple quantum convolutional gates, and a quantum feature map is generated through quantum circuit convolution operation on the input quantum state.

[0054] Step 4: The quantum feature map extracted by the quantum convolutional layer is input into a quantum feature fusion module. The module uses quantum entanglement and interference characteristics to fuse quantum features of different levels and scales by designing quantum circuits, and generates more abstract and advanced quantum feature representations.

[0055] Step 5: The quantum state output by the quantum feature fusion module is converted into classical data through a measurement operation and input into a fully connected layer.

[0056] Step 6: The trained model is tested and predicted, and the model prediction result is output.

[0057] The data enhancement specifically includes: randomly rotating the input image data, with the rotation angle randomly selected between 0 and 360 degrees; performing Gaussian blur processing on the rotated image data, with the Gaussian kernel size randomly selected, to enhance the diversity of quantum data; and performing occlusion at random positions of the image data, with the occluded area accounting for 5% to 15% of the entire image data, and the pixel values of the occluded area being randomly generated. Through the above data enhancement operations, the generalization ability and robustness of the quantum convolutional neural network can be effectively improved, so that it can better adapt to different image data distributions.

[0058] The classical convolution layer specifically includes: performing convolution operation on the enhanced classical data using a convolution kernel to extract local features; then introducing a non-linear factor through an activation function to enhance the expression ability of the network; and preparing for the subsequent quantum convolution layer.

[0059] C = f(X⊙W1 + b1) = r(X⊙W1 + b1)

[0060] In the above formula, C is the output of the convolution layer; W1 is the weight matrix, b1 is the bias matrix; r is the ReLU activation function; ⊙ is the convolution operation, and X is the input enhanced classical data.

[0061] The specific steps of the quantum convolution layer are as follows:

[0062] 1) Map the local features output by the classical convolution layer to the quantum state space through the encoding module E. This module uses single-bit quantum rotation gate R y (θ) and Hadamard gate H to encode the input data into quantum states for processing in the quantum circuit.

[0063]

[0064]

[0065] The input data of the quantum convolution layer is used as the rotation angle θ of R y (θ).

[0066] 2) Transform and feature extract the encoded quantum state using the variational quantum circuit module U. This module is composed of a series of parameterized quantum gates, including parameterized two-qubit controlled gate CR z (φ). By optimizing the parameters of these gates, the variational quantum circuit can learn the latent feature representation of the data.

[0067]

[0068] For the quantum state encoded by the encoding module E, CR z(φ) enables rotating the target qubit by an angle φ around the Z axis if the control qubit is |1>. φ is a learnable parameter.

[0069] 3) The measurement module M measures the quantum state output by the variational quantum circuit and converts the quantum state into classical information. The measurement result is used as the quantum feature map output by the quantum convolution layer and passed to the subsequent quantum feature fusion module for further processing.

[0070]

[0071] The measurement module M uses the Pauli-Z gate Z to measure each target quantum state.

[0072] The quantum feature fusion module is specifically as follows: multiple quantum features extracted by the quantum convolution layer are mapped to the high-dimensional Hilbert space, and the probability amplitude of each quantum feature is linearly combined using the quantum superposition principle to construct a superposition state, thereby realizing the fusion of quantum features in high-dimensional space. At the same time, through quantum entanglement operations, the quantum bits of different quantum features are entangled, non-local correlations between quantum features are established, and long-range dependencies between features are captured. The combination of quantum superposition and quantum entanglement gives full play to the unique advantages of quantum mechanics and realizes the efficient fusion of quantum features. The fusion coefficient formula is expressed as:

[0073] f′1(w1)=f1(w1)+a1f1(w1),

[0074] f′2(w2)=f2(w2)+a2f2(w2),

[0075] f′3(w3)=f3(w3)+a3f3(w3),

[0076] f′4(w4)=f4(w4)+a4f4(w4),

[0077] f′(W)=(a1f1(w1)+a2f2(w2)+a3f3(w3)+a4f4(w4)) / 4,

[0078]

[0079] Among them, f i (w i ) is the feature map of layer i, w i is the quantum convolution layer generating f i (w i ) The weight used, a i The f generated by the feature fusion module i (w i ) weight coefficient, f′ i (w i ) is in accordance with ai The i-th layer feature map after weighted fusion. f'(W) is the new feature map obtained after adding f'(W) i (w i ) together. is the new feature map obtained by channel superposition fusion of f'(W), f'(W) i (w i ).

[0080] The specific steps of the full connection layer are:

[0081] C F = f(Q⊙W2+b2) = σ(Q⊙W2+b2)

[0082] In the above formula, C F is the full connection layer, Q is the output of the quantum feature fusion module; W1 is the weight matrix, b1 is the bias matrix; σ is the ReLU activation function; and is the convolution operation.

[0083] The training of the quantum convolutional neural network is completed through the above steps. When identifying and classifying for computer vision, the images collected by the vision sensor in real time are input into the trained quantum convolutional neural network, and the identification and classification of the target in the image are obtained.

[0084] Example Analysis 1:

[0085] Step 1: Data augmentation: In order to improve the generalization ability and robustness of the model, we performed data augmentation on the MNIST dataset. The specific operations include: random rotation (-10°-10°), random translation (-22 pixels), random scaling (0.9-1.1 times), etc. The augmented dataset is expanded to 5 times the original dataset.

[0086] Step 2: Classical convolutional layer: We first use two classical convolutional layers to extract features from the augmented image data. The first convolutional layer contains 16 3x3 convolutional kernels with a stride of 1 and padding of 1; the second convolutional layer contains 32 3x3 convolutional kernels with a stride of 1 and padding of 1. After each convolutional layer, a 2x2 max pooling layer and a ReLU activation function are connected.

[0087] Step 3: Quantum convolutional layer: After the classical convolutional layer, we introduce a quantum convolutional layer to further extract quantum features of the image. The quantum convolutional layer is composed of multiple quantum bits, each representing a quantum feature. We use 4 quantum bits to evolve the quantum state through a series of quantum gate operations (such as Hadamard gate, CNOT gate, etc.), extracting the quantum features of the image.

[0088] Step 4 Quantum Feature Fusion Module: To fully utilize the features extracted by both classical and quantum convolution layers, we designed a quantum feature fusion module. This module uses an attention mechanism to weight the fusion of classical and quantum features based on their importance. The fused feature vector contains both local and global information of the image.

[0089] Step 5 Fully Connected Layers: The fused feature vector is passed through two fully connected layers for classification prediction. The first fully connected layer contains 128 neurons with ReLU activation function, and the second fully connected layer contains 10 neurons corresponding to the 10 classes of the MNIST dataset, with a Softmax activation function outputting classification probabilities.

[0090] Step 6 Model Training: We trained the model using the augmented MNIST training set, with the cross-entropy loss function as the optimization objective. We used the Adam optimizer with an initial learning rate of 0.001, a batch size of 128, and trained for 100 epochs. During training, we used early stopping and learning rate decay strategies to prevent overfitting.

[0091] Step 7 Model Prediction: We evaluated the trained model's performance on the MNIST test set. Our model achieved an accuracy of 92.1% on the test set, surpassing traditional convolutional neural network models. This indicates that the quantum feature fusion module effectively extracts and fuses classical and quantum features of images, improving the model's classification performance.

[0092] The experimental results are shown in Figure 2 、 Figure 3 .

[0093] Step 1 Data Augmentation: Before training the model, we performed data augmentation on the Cifar-10 dataset. This includes random horizontal flipping, random cropping, and standardization. Through data augmentation, we can effectively expand the training samples, improve the model's generalization ability, and reduce the risk of overfitting.

[0094] Step 2 Classical Convolution Layers: In the quantum convolutional neural network, we first use classical convolution layers to extract local features of images. Classical convolution layers consist of multiple convolution kernels that can effectively capture local features such as edges and textures in images. In this example, we use two classical convolution layers with kernel sizes of 3x3 and 5x5, and ReLU activation functions.

[0095] Step 3 Quantum Convolutional Layer: After the classical convolutional layer, we introduce a quantum convolutional layer to further extract quantum features of the image. The quantum convolutional layer utilizes quantum circuits to implement convolution operations, and through the combination and parameterization of quantum gates, it can generate more rich and expressive features. In this example, we use a quantum convolutional layer with a depth of 4 and use 6 qubits.

[0096] Step 4 Quantum Feature Fusion Module: To make full use of the features extracted by the classical convolutional layer and the quantum convolutional layer, we design a quantum feature fusion module. This module fuses classical features and quantum features through feature concatenation and attention mechanisms to generate more complete feature representations. Specifically, we concatenate the classical features and quantum features in the channel dimension, and then use a self-attention mechanism to adaptively adjust the importance of the features, finally obtaining the fused features.

[0097] Step 5 Fully Connected Layer: After feature fusion, we use a fully connected layer for classification prediction. The fully connected layer can map high-dimensional features to class label space, realizing the task of image classification. In this example, we use two fully connected layers, using ReLU activation function in the middle, and using Softmax function in the last layer for multi-classification.

[0098] Step 6 Model Training: During model training, we use the cross-entropy loss function to measure the difference between the predicted results and the true labels, and use the Adam optimizer to update the model parameters. We divide the dataset into training set and validation set, and evaluate the model performance on the validation set after each epoch to monitor the training process and prevent overfitting.

[0099] Step 7 Model Prediction: After training, we use the trained model to predict the test set. For each test sample, we input it into the model and get the prediction result through forward propagation. We use Top-1 accuracy to evaluate the performance of the model, that is, the proportion of samples whose predicted results completely match the true labels. On the Cifar-10 dataset, our quantum convolutional neural network method achieves high accuracy, demonstrating the effectiveness of quantum feature fusion.

[0100] The experimental results are shown in Figure 4 , Figure 5 .

Claims

1. A method for improving quantum feature fusion of a quantum convolutional neural network, characterized in that, The method comprises the following steps: (1) performing enhancement processing on the input training data set to generate a plurality of sample data after data volume expansion; (2) inputting the preprocessed sample data into a convolutional neural network to extract local features of the image and generate a feature map; (3) converting the feature map into a quantum state, inputting the quantum state into a quantum convolutional layer, performing convolution operation on the input quantum state through a quantum circuit, and generating a quantum feature map; (4) inputting the quantum feature map extracted by the quantum convolutional layer into a quantum feature fusion module to fuse quantum features of different levels and scales and generate a quantum feature representation; (5) inputting the quantum feature representation output by the quantum feature fusion module into a fully connected layer through a measurement operation to map the quantum feature to an output space and obtain a prediction result.

2. The method for improving quantum feature fusion of a quantum convolutional neural network according to claim 1, wherein, The method comprises the following steps: randomly rotating the input image data, and randomly selecting a rotation angle between 0 and 360 degrees; performing Gaussian blur processing on the rotated image data; occluding the image data at a random position, the occluded area accounting for 5% to 15% of the entire image data, and the pixel value of the occluded area being randomly generated.

3. The method of claim 1, wherein, The method comprises the following steps: 1) Map feature map data to quantum state space by encoding module E; encode input data as quantum states using rotation gate R y (θ) and Hadamard gate H; 2) using a variational quantum circuit module U to transform and extract features of the encoded quantum state; 3) a measurement module M measures the quantum state output by the variational quantum circuit module U, converts the quantum state, and outputs a quantum feature map.

4. The method of claim 1, wherein, The method comprises the following steps: mapping a plurality of quantum features of the quantum feature map to a Hilbert space, linearly combining the probability amplitudes of each quantum feature by using the quantum superposition principle, constructing a superposition state, and realizing fusion of the quantum features in the Hilbert space; The fusion coefficient formula is represented as: f1'(w1) = f1(w1) + a1f1(w1), f2'(w2) = f2(w2) + a2f2(w2), f3'(w3) = f3(w3) + a3f3(w3), f4'(w4) = f4(w4) + a4f4(w4), f'(W) = (a1f1(w1) + a2f2(w2) + a3f3(w3) + a4f4(w4)) / 4, wherein f i (w i ) is the i-th layer quantum feature map, w i is the weight used by the quantum convolution layer to generate the quantum feature map f i (w i ), a i is the weight coefficient of f i (w i ), f i ′(w i ) is the i-th layer feature map after weighted fusion according to a i ; f′(W) is a new feature map obtained after adding f i ′(w i ) for fusion, W = {w1, w2, w3, w4}; is a new feature map obtained after channel superposition fusion of f′(W) and f i ′(w i ), i.e., a quantum feature representation.

5. The method of claim 1, wherein, The method comprises the following steps: C F = f(Q ® W2+ b2) = σ(Q ® W2+ b2) C in the above formula F is a fully connected layer, and Q is a quantum feature representation output by the quantum feature fusion module. W2 is a weight matrix, b2 is a bias matrix, σ is a ReLU activation function, and is a convolution operation.

6. A system for improving quantum feature fusion of a quantum convolutional neural network, the system comprising: The method comprises the following steps: a data enhancement module for performing enhancement processing on the input training data set to generate a plurality of sample data after data volume expansion; a convolutional neural network module for inputting the preprocessed sample data into a convolutional neural network to extract local features of the image and generate a feature map; a quantum convolution layer, configured to convert a feature map into a quantum state, and perform a convolution operation on the input quantum state through a quantum circuit to generate a quantum feature map; a quantum feature fusion module, configured to input the quantum feature map extracted by the quantum convolution layer into the quantum feature fusion module, fuse quantum features of different levels and scales, and generate a quantum feature representation; a fully connected layer, configured to input the quantum feature representation output by the quantum feature fusion module into the fully connected layer through a measurement operation, map the quantum feature to an output space, and obtain a prediction result.

7. An apparatus for improving quantum feature fusion of a quantum convolutional neural network, the apparatus comprising: The method comprises a memory and a processor; the memory is configured to store a computer program; the processor is configured to implement the method for quantum feature fusion of the improved quantum convolutional neural network according to any one of claims 1-5 when the computer program is executed.

8. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, and when the computer program is executed by the processor, the method for quantum feature fusion of the improved quantum convolutional neural network according to any one of claims 1-5 is implemented.

Images