改进生成式对抗网络的图像数据集增强算法

针对现有深度学习中图像数据集缺乏的问题，提出了一种基于深度卷积生成式对抗网络(Deep Convolutional Generative Adversarial Network, DCGAN)的图像数据集增强算法。该算法对DCGAN网络进行改进，首先在不过多增加计算量的前提下改进现有的激活函数，增强生成特征的丰富性与多样性；然后通过引入相对判别器有效缓解模式坍塌现象，从而提升模型稳定性；最后在现有生成器结构中引入残差块，获得相对高分辨率的生成图像。实验结果表明，将所提方法应用在MNIST、SAR和医学血细胞数据集上，图像数据增强效果与未改进的DCGAN网络相比显著提升。

An improved image dataset enhancement algorithm for generative adversarial network

For the problem of lacking image datasets in existing deep learning,an image dataset augmentation optimization algorithm based on deep convolutional generative adversarial network(DCGAN) is proposed.The algorithm improves the DCGAN network.First,the existing activation function is improved on the premise of not increasing the amount of computation to enhance the richness and diversity of generated characteristics.Then,a relativistic discriminator is introduced to alleviate the collapse of the model effectively and improve model stability.Finally,the residual block is introduced into the existing network structure to obtain a relatively high-resolution generated image.The experimental results show that,compared with the unimproved DCGAN,the proposed algorithm in this paper has better image data augmentation effect on MNIST,SAR and medical blood cell datasets.