改进生成对抗网络在图片数据生成中的应用

doi:10.3778/j.issn.1002-8331.2012-0429

摘要/Abstract

摘要： 图片数据生成旨在根据现有的图片数据，产生与原始图片数据分布相似的图片数据。当前主流的生成对抗网络模型（generative adversarial networks，GAN）产生的图片数据质量较差，模型的训练总是遇到调试困难、训练不稳定、梯度消失、模式崩溃等一系列问题。根据稀疏表达结构和残差结构组合而成的生成器，残差结构组成的辨别器，提出了一种能够生成高质量图片的GAN模型。根据分支网络模型构成的生成器，设计了多种类图片数据生成模型，可以使用一个模型同时训练生成多种类型的图片数据。为了更好地对数据进行训练，设计了一种动态匀速下降学习率，能够根据运行时间对学习率的衰减进行指导。在各个数据集上的实验结果表明，改进模型结构在图像数据生成上比其他算法更加稳定、鲁棒，能够生成更高质量的图片数据。

关键词: 图片数据生成, 生成对抗网络（GAN）, 残差结构, 动态学习率

Abstract: Image data generation aims to generate image data with similar distribution to original image data according to existing image data. At present, the image data quality produced by the mainstream generative adversarial networks（GAN） model is poor, and the training of the model always encounters a series of problems such as debugging difficulties, unstable training, gradient disappearance, mode collapse, and so on. According to the generator composed of sparse expression structure and residual structure, and the discriminator composed of residual structure, a GAN model capable of generating high-quality pictures is proposed. According to the generator composed of the branch network model, a generation model of various kinds of picture data is designed, and one model can be used to train and generate various types of picture data at the same time. To train the data better, a dynamic learning rate with uniform decline is designed, which can guide the attenuation of the learning rate according to the running time. Experimental results on various data sets show that the improved model structure is more stable and robust than other algorithms in image data generation, and can generate higher quality image data.

Key words: image data generation, generative adversarial networks（GAN）, residual structure, dynamic learning rate

孟辰, 曾志高, 朱艳辉, 朱文球, 易胜秋, 董丽君. 改进生成对抗网络在图片数据生成中的应用[J]. 计算机工程与应用, 2022, 58(15): 260-269.

MENG Chen, ZENG Zhigao, ZHU Yanhui, ZHU Wenqiu, YI Shengqiu, DONG Lijun. Application of Improved Generative Adversarial Networks in Image Data Generation[J]. Computer Engineering and Applications, 2022, 58(15): 260-269.

参考文献

[1] 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.
LIANG J J，WEI J J，JIANG Z F.Generative adversarial networks GAN overview[J].Journal of Frontiers of Computer Science and Technology，2020，14（1）：1-17
[2] MAHALINGAIAH K，MATICHUK B.ProDeblurGAN：progressive growing of GANs for blind motion deblurring in face recognition[C]//International Conference on Smart Multimedia.Cham：Springer，2019：439-450.
[3] KARRAS T，LAINE S，AITTALA M，et al.Analyzing and improving the image quality of stylegan[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition，2020：8110-8119.
[4] HUANG X，LI Y，POURSAEED O，et al.Stacked generative adversarial networks[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition，2017：5077-5086.
[5] OLIEHOEK F A，SAVANI R，GALLEGO J，et al.Beyond local Nash equilibria for adversarial networks[C]//Benelux Conference on Artificial Intelligence.Cham：Springer，2018：73-89.
[6] RADFORD A，METZ L，CHINTALA S.Unsupervised representation learning with deep convolutional generative adversarial networks[J].arXiv：1511.06434，2015.
[7] GULRAJANI I，AHMED F，ARJOVSKY M，et al.Improved training of wasserstein gans[J].Advances in Neural Information Processing Systems，2017，30：5767-5777.
[8] WEI X，GONG B，LIU Z，et al.Improving the improved training of wasserstein gans：a consistency term and its dual effect[J].arXiv：1803.01541，2018.
[9] BARRATT S，SHARMA R.A note on the inception score[J].arXiv：1801.01973，2018.
[10] WU Z，SHEN C，VAN DEN HENGEL A.Wider or deeper：revisiting the resnet model for visual recognition[J].Pattern Recognition，2019，90：119-133.
[11] SAGE D，DONATI L，SOULEZ F，et al.DeconvolutionLab2：an open-source software for deconvolution microscopy[J].Methods，2017，115：28-41.
[12] SANTURKAR S，TSIPRAS D，ILYAS A，et al.How does batch normalization help optimization?[J].Advances in Neural Information Processing Systems，2018，31：2483-2493.
[13] DUBEY A K，JAIN V.Comparative study of convolution neural network’s relu and leaky-relu activation functions[M]//Applications of computing，automation and wireless systems in electrical engineering.Singapore：Springer，2019：873-880.
[14] ZHANG D，LI G Z，WANG H G，et al.Improvement of adaptive learning rate for pose estimation[J].Journal of Electronic Measurement and Instrumentation，2019.
[15] S?NDERBY C K，CABALLERO J，THEIS L，et al.Amortised map inference for image super-resolution[J].arXiv：1610.04490，2016.
[16] SHI W Z，CABALLERO J，HUSZáR F，et al.Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network[J].arXiv：1609.05158，2016.
[17] HEUSEL M，RAMSAUER H，UNTERTHINER T，et al.Gans trained by a two time-scale update rule converge to a local nash equilibrium[J].arXiv：1706.08500，2017.
[18] 卢倩雯，陶青川，赵娅琳，等.基于生成对抗网络的漫画草稿图简化[J].自动化学报，2018，44（5）：840-854.
LU Q W，TAO Q C，ZHAO Y L.Sketch simplification using generative adversarial networks[J].Acta Automatica Sinica，2018，44（5）：840-854
[19] GUO T，XU C，HUANG J，et al.On positive-unlabeled classification in GAN[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition，2020：8385-8393.
[20] SRIVASTAVA N，HINTON G，KRIZHEVSKY A，et al.Dropout：a simple way to prevent neural networks from overfitting[J].Journal of Machine Learning Research，2014，15（1）：1929-1958.
[21] XU Q，HUANG G，YUAN Y，et al.An empirical study on evaluation metrics of generative adversarial networks[J].arXiv：1806.07755，2018.