在深度学习中，数据的体量和质量是影响模型性能的重要因素。深度卷积生成对抗网络作为一种新型无监督模型，采用生成器和判别器的对抗学习思想生成新的图像数据集，解决了传统数据增强方法无法提取更多图像细节的缺陷，但存在生成图像质量较差、模型不稳定的问题。针对以上问题，本文从外部输入噪声维数和内部结构两个方面对深度卷积生成对抗网络模型进行改进，提出一种基于相对判别器的深度残差卷积生成对抗网络模型。主要工作如下：
针对最大似然维数算法估计图像本征维数存在的负偏差现象，提出采用自适应最大似然维数估计算法估计图像本征维数。通过对最大似然估计算法求出的本征维数进行加权求和后再取平均值，削弱了不相干数据点的贡献，加强了重要区域数据点的作用，并根据其结果确定深度卷积生成对抗网络的最佳噪声输入维数。实验结果表明，使用改进的最大似然维数估计算法进行图像本征维数估计，可以减少模型的计算量，提高模型的生成图像质量。
针对深度卷积生成对抗网络存在生成图像质量差、模型崩塌的问题，提出一种基于相对判别器的深度残差卷积生成对抗网络模型。首先，采用SeLU激活函数和相对判别器作为生成对抗网络的判别器结构，增强生成图像的质量与多样性。其次，在现有生成器结构中引入残差块，在提升模型捕获图像细节特征能力的同时提高了模型的稳定性。通过在MNIST、fashion-MNIST和MSTAR数据集上进行实验仿真，结果表明，相比于深度卷积生成对抗网络，本文改进算法在三种数据集上的FID值分别下降29.60%、18.71%和1.90%，图像数据增强效果显著提升。

The volume and quality of data are important factors that influence model performance in deep learning. As a new type of unsupervised model, Deep Convolutional Generative Adversarial Network uses the adversarial learning idea of generator and discriminator to generate new image datasets, which solves the problem that traditional data enhancement methods cannot extract more image details, but this model has the problems of poor image quality and unstable model. In view of the above problems, this thesis improves the Deep Convolutional Generative Adversarial Network model from two aspects of external input noise dimension and internal structure, and proposes a Relativistic and Residual Deep Convolutional Generative Adversarial Network. The main work is as follows:

The adaptive maximum likelihood dimension estimation algorithm is proposed to estimate the image intrinsic dimension in response to the negative bias phenomenon in the estimation of the image intrinsic dimension by the maximum likelihood dimension algorithm. By weighting and summing the intrinsic dimensions obtained by the maximum likelihood estimation algorithm and then taking the average value, the contribution of irrelevant data points is weakened, and the role of data points in important regions is strengthened. The optimal noise input dimension of the network is determined based on the results. The experimental results show that the use of the improved maximum likelihood dimension estimation algorithm to estimate the intrinsic dimension of the image can reduce the calculation amount of the model and improve the generation effect of the model.

A Relativistic and Residual Deep Convolutional Generative Adversarial Network is proposed to address the problems of poor image quality and model collapse. Firstly, the SeLU activation function and the relative discriminator are used as the discriminator structure of the Generative Adversarial Network to enhance the quality and diversity of the generated images. And then, a residual block is introduced into the existing generator, which improves the ability of the model to capture image detail features while improving the stability of the model. Through experimental simulations on the MNIST, fashion-MNIST and MSTAR datasets, the results show that, compared with the Deep Convolutional Generative Adversarial Network, the FID of the improved algorithm in this thesis on the three datasets are reduced by 29.60%, 18.71%, 1.90% respectively, and the image data enhancement effect is significantly improved.

[1]张蔚敏, 蒋阿芳, 纪学毅. 人工智能芯片产业现状[J]. 电信网技术, 2018(2):67-71.

[2]王哲. 2021年中国人工智能产业发展形势展望[R]. 北京: 国家工业和信息化部中国电子信息产业发展研究院, 2021.

[3]王坤峰, 苟超, 段艳杰, 等. 生成式对抗网络GAN的研究进展与展望[J]. 自动化学报, 2017, 43(3): 321-332.

[4]Taigman Y, Yang M, Ranzato M A, et al. Deepface:Closing the gap to human-level performance in face verification[C]// Proceedings of the IEEE conference on computer vision and pattern recognition. Columbus: IEEE, 2014: 1701-1708.

[5]Lin T Y, Maire M, Belongie S, et al. Microsoft COCO: Common Objects in Context[C]// European Conference on Computer Vision. Zurich: ECCV, 2014: 740-755.

[6]Simonyan K, Zisserman A. VGGNet-Very Deep Convolutional Networks for Large-Scale Image Recognition[EB/OL]. arXiv preprint arXiv:1409.1556, 2014-09.

[7]He K, Zhang X, Ren S, et al. Deep residual learning for image recognition[C]// Proceedings of IEEE Conference on Computer Vision and Pattern Recognition. Washington DC: IEEE 2016: 770-778.

[8]Krizhevsky A, Sutskever I, Hinton G E. Imagenet classification with deep convolutional neural networks[C]// Advances in neural information processing systems 25. Cambridge: MIT Press, 2012: 1097-1105.

[9]Goodfellow I, Pouget-Abadie J, Mirza M, et al. Generative adversarial nets [C]// Advances in neural information processing systems 27. mount royal: MIT Press, 2014:2672-2680.

[10]Mirza M, Osindero S. Conditional generative adversarial nets[J] Computer Science, 2014, 11(1): 2672-2680.

[11]Radford A, Metz L, Chintala S. Unsupervised representation learning with deep convolutional generative adversarial networks[EB/OL]. https://arXiv.org/abs/1511.06434, 2016-01.

[12]Mao X, Li Q, Xie H, et al. Least squares generative adversarial networks[C]// IEEE Conference on Computer Vision and Pattern Recognition. Venice: IEEE, 2017: 2794-2802.

[13]Arjovsky M, Chintala S, Bottou L. Wasserstein generative adversarial networks[C]// International Conference on Machine Learning. New York: ACM, 2017: 214-223.

[14]Miyato T, Kataoka T, Koyama M, et al. Spectral normalization for generative adversarial networks[C]// International Conference on Learning Representations. Vancouver: MIT Press, 2018: 1539-1542.

[15]吴少乾, 李西明. 生成对抗网络的研究进展综述[J]. 计算机科学与探索, 2020, 14(3): 377-388.

[16]Choi Y, Choi M, Kim M, et al. StarGAN: Unified Generative Adversarial Networks for Multi-Domain Image-to-Image Translation[C]// Proceedings of the 2018 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: CVPR, 2018: 8789-8797.

[17]Brock A, Donahue J, Simonyan K. Large scale GAN training for high fidelity natural image synthesis[C]// International Conference on Learning Representations. New Orleans, ICLR, 2019: 1-29.

[18]梁俊杰, 韦舰晶, 蒋正锋. 生成对抗网络GAN综述[J]. 计算机科学与探索, 2020, 14(1): 1-17.

[19]Öcal A, Özbakır L. Supervised deep convolutional generative adversarial networks[J]. Neurocomputing, 2021, 449(8): 389-398.

[20]代亮, 梅洋, 李曙光, 等. 基于对称残差U型网络的路网交通流量数据修复[J]. 交通运输系统工程与信息, 2020, 20(5): 93-99.

[21]黄淑英, 汪斌, 李红霞, 等. 基于生成对抗网络的图像去雾算法[J]. 模式识别与人工智能, 2021, 34(11): 990-1003.

[22]全海燕, 王涛, 郑志清. 加性频域分解的生成对抗网络语音去混响[J/OL]. 工程科学与技术, 2022-03-10.

[23]张晓峰, 吴刚. 基于生成对抗网络的数据增强方法[J]. 计算机系统应用, 2019, 28(10): 201-206.

[24]王超学, 张涛, 马春森. 面向不平衡数据集的改进型SMOTE算法[J]. 计算机科学与探索, 2014, 8(6): 91-98.

[25]黎旭, 陈家兑, 吴永明, 等. 基于改进SMOTE的制造过程不平衡数据分类策略[J/OL]. 计算机工程与应用, 2022-01-28.

[26]梅大成, 陈江, 郑涛. 边界与密度适应的SMOTE算法研究[J/OL]. 计算机应用研究, 2022-03-19.

[27]Cubuk E D, Zoph B, Mane D, et al. Auto Augment: Learning Augmentation Strategies From Data[C]// The IEEE Conference on Computer Vision and Pattern Recognition. Long Beach, CVPR, 2019: 113-123.

[28]Söderman M, Ikander P, Boljanovic S, et al. Utilizing the lateral excess for autologous augmentation in massive weight loss patients[J]. Gland Surgery, 2019,8(04): 271-275.

[29]Naghizadeh A, Abavisani M, Metaxas D N. Greedy autoaugment[J]. Pattern Recognition Letters, 2020, 138(8): 624-630.

[30]陈佛计, 朱枫, 吴清潇, 等. 生成对抗网络及其在图像生成中的应用研究综述[J]. 计算机学报, 2021, 44(02): 347-369.

[31]Rafael A Z, Esther L C. Parkinson’s Disease EMG Data Augmentation and Simulation with DCGANs and Style Transfer[J]. Sensors, 2020, 09(20): 2605-2628.

[32]Song Y, Li Y, Wang Y, Wang Y. Data Augmentation for Imbalanced HRRP Recognition Using Deep Convolutional Generative Adversarial Network[J], IEEE ACCESS, 2020, 10(8): 201686-201695.

[33]Ahn G, Choi B S, Ko S, et al. High-Resolution Knee Plain Radiography Image Synthesis Using Style Generative Adversarial Network Adaptive Discriminator Augmentation[J]. Journal of Orthopaedic Surgery and Research, 2022, 3(1): 1-26.

[34]甘岚, 沈鸿飞, 王瑶, 等. 基于改进DCGAN的数据增强方法[J]. 计算机应用, 2021, 41(5): 1305-1313.

[35]裴卉宁, 谭昭芸, 张金勇, 等. DCGAN在汽车造型设计模型中的应用[J/OL]. 机械科学与技术, 2021-07-12.

[36]丁斌, 夏雪, 梁雪峰. 基于深度生成对抗网络的海杂波数据增强方法[J]. 电子与信息学报, 2021, 43(7): 1985-1991.

[37]Yinka-Banjo C, Ugot O A. A review of generative adversarial networks and its application in cybersecurity[J]. Artificial Intelligence Review, 2020, 53(3): 1721-1736.

[38]Abolhasannejad V, Huang X M, Namazi N, et al. Developing an optical image-based method for bridge deformation measurement considering camera motion[J]. Sensors, 2018, 18(9): 1-18.

[39]杨毅, 卢诚波, 徐根海. 面向不平衡数据集的一种精化Borderline-SMOTE方法[J]. 复旦学报(自然科学版), 2017, 56(05): 537-544.

[40]Naghizadeh A, Metaxas D N, Liu D. Greedy auto-augmentation for n-shot learning using deep neural networks[J]. Neural Networks, 2020, 135(11): 68-77.

[41]Klambauer G, Unterthiner T, Mayr A, et al. Self-normalizing neural networks[C]// Proceeding of the 31th International Conference on Neural Information Processing Systems 30. Long Beach: MIT Press, 2017: 971-980.

[42]魏富强, 古兰拜尔·吐尔洪, 买日旦·吾守尔. 生成对抗网络及其应用研究综述[J]. 计算机工程与应用, 2021, 57(19): 18-31.

[43]胡龙辉, 王朝立, 孙占全, 等. 基于WGAN的图像识别方法[J]. 控制工程, 2020, 27(12): 2168-2175.

[44]段雪源, 付钰, 王坤. 基于VAE-WGAN的多维时间序列异常检测方法[J/OL]. 通信学报: 2022-03-20.

[45]张哲新, 原俊青, 郭欢磊, 等.多判别器协同框架:高品质图像的谱归一生成对抗网络[J].小型微型计算机系统, 2021, 42(1): 201-207.

[46]杨明. 面向分类的高光谱影像特征提取技术研究[D]. 解放军信息工程大学, 2012.

[47]张荣国, 姚晓玲, 赵建, 等. 融入局部几何特征的流形谱聚类图像分割[J]. 模式识别与人工智能, 2020, 33(04): 313-324.

[48]Camastra F, Staiano A. Intrinsic dimension estimation: Advances and open problems[J]. Information Sciences, 2016, 328(01): 26-41.

[49]Halimi A, Honeine P, Kharouf M, et al. Estimating the Intrinsic Dimension of Hyperspectral Images Using a Noise-Whitened Eigengap Approach[J]. IEEE Transactions on Geoscience and Remote Sensing, 2016, 54(7): 3811-3821.

[50]郭伟, 庞晨. 改进生成式对抗网络的图像数据集增强算法[J]. 电讯技术: 2022, 62, (03): 281-287.

[51]Jolicoeur-Martineau A. The relativistic discriminator: a key element missing from standard GAN[EB/OL]. arXiv preprint arXiv:1807.00734, 2018-09.

[52]王海文. 基于生成式对抗网络的数据增强方法研究[D]. 南京: 南京邮电大学, 2019.

[53]Simon H. 神经网络原理[M]//叶世伟. 北京: 机械工业出版社, 2004: 81-175.