【作者】 陈超；

【导师】 王允乾； 刘韬；

【作者基本信息】 南昌大学 ， 电子信息（专业学位）， 2024， 硕士

【摘要】 图像分类是人工视觉系统中的一个重要课题,在过去的一段时间里引起了人们的极大兴趣和广泛关注。该领域旨在根据图像的视觉内容对输入图像进行分类,但是,面对数据量爆炸性的增长、图像信息的复杂度提高等现状时,经典机器学习模型在分类速度、准确性和泛化能力等指标上开始显得捉襟见肘。随着量子信息技术从理论到现实落地,量子计算领域的兴起,量子机器学习成为一个炙手可热的研究方向。它利用量子计算中的叠加、纠缠、并行等特性,在特定任务中表现优异。在NISQ时代下,量子器件的规模限制与图像尺寸之间的矛盾,使得研究人员开始将目光投向结合二者的优势,开发出颇具潜力的混合经典-量子神经网络模型。为了打破经典计算与量子计算的壁垒,突破量子机器学习模型目前的输入局限性,本文基于参数化量子线路和迁移学习、知识蒸馏,设计并实现了量子特征处理器,构建了混合经典-量子图像分类模型,具体研究内容如下:首先,本文提出了一种用于图像分类任务的量子特征处理器,结合了传统卷积神经网络和量子计算的优势,旨在提高图像分类的准确性和效率。首先使用经过预训练的经典Dense Net121对输入图像进行特征提取,我们将这些特征以旋转角编码到量子比特中,并利用量子门设计量子线路,对特征进行转换和融合,最后对量子比特进行测量,以获得最终的分类结果。我们设计了多个实验,在COVID-19放射学数据集上的实验结果表明,量子特征处理器具有处理图像特征的能力,加入量子特征提取器能够有效提高经典模型的特征处理能力,所提出的混合经典-量子模型在分类任务上相比于现有研究成果获得了更优秀的分类性能。其次,我们结合知识蒸馏和量子特征处理器,利用知识蒸馏技术,实现将知识从经典神经网络模型中转移到量子神经网络模型中,有效填补了经典机器学习模型和量子机器学习之间的差距,使得小型量子神经网络模型成为可能。我们使用Res Net-18作为教师模型,利用KL散度计算后的Loss作为反向传播过程中的监督信号,递交给学生量子模型,从而促进学生模型的训练。通过实验结果验证,经典深度网络模型中的知识可以有效地通过蒸馏递交给学生网络模型,并且所提出的学生模型在图像分类任务上具有鲁棒性。更多还原

【Abstract】 Image classification is an important topic in artificial vision systems,which has attracted great interest and widespread attention in the past period.The field aims to classify input images based on their visual content.However,classical machine learning models are beginning to show limitations in terms of classification speed,accuracy,and generalization ability,especially in the face of explosive growth in data volume and increased complexity of image information.With the implementation of quantum information technology from theory to reality,and the rise of quantum computing,quantum machine learning has become a hot research direction.It leverages the properties of quantum computing such as superposition,entanglement,and parallelism to perform exceptionally well in specific tasks.In the era of Noisy Intermediate-Scale Quantum(NISQ)devices,the contradiction between the scale limitations of quantum devices and the size of images has led researchers to focus on combining the advantages of both to develop a promising hybrid classical-quantum neural network model.In order to break the barriers between classical and quantum computing and overcome the current input limitations of quantum machine learning models,this paper designs and implements a quantum feature processor,proposes a hybrid classical-quantum image classification model based on parameterized quantum circuits,transfer learning,and knowledge distillation.The specific research content is as follows:Firstly,this paper proposes a quantum feature processor for image classification tasks,which combines the advantages of traditional convolutional neural networks and quantum computing,aiming to improve the accuracy and efficiency of image classification.Initially,pre-trained classical Dense Net121 is used for feature extraction from input images.These features are then encoded into quantum bits with rotation angle encoding,and quantum gates are utilized to design quantum circuits for feature transformation and fusion.Finally,the qubits are measured to obtain the final classification results.Multiple experiments conducted on the COVID-19 Radiography Dataset demonstrate that the quantum feature processor has the ability to process image features,and integrating the quantum feature extractor can effectively enhance the feature processing capability of classical models.The proposed hybrid classicalquantum model achieves better classification performance in classification tasks compared with existing research results.Then,we combine knowledge distillation with the quantum feature processor,utilizing knowledge distillation techniques to transfer knowledge from classical neural network models to quantum neural network models.This effectively bridges the gap between classical machine learning models and quantum machine learning,enabling the development of small-scale quantum neural network models.We employ Res Net-18 as teacher models,using the loss calculated by KL divergence as a supervision signal in the backpropagation process,which is then passed to the student quantum model to facilitate its training.Experimental results validate that the knowledge in classical deep network models can be distilled to the student network model effectively,and the proposed student model demonstrates robustness in image classification tasks.更多还原