Abstract:

Concerning the lack of flexibility in adversarial training of Deep Convolutional Generative Adversarial Network （DCGAN） and the problems of inflexible optimization and unclear convergence state of Binary Cross-Entropy loss （BCE loss） function used in DCGAN， an improved algorithm of Generative Adversarial Network （GAN） based on arbitration mechanism was proposed. In this algorithm， the proposed arbitration mechanism was added on the basis of DCGAN. Firstly， the network structure of the proposed improved algorithm was composed of generator， discriminator， and arbiter. Secondly， the adversarial training was conducted by the generator and discriminator according to the training plan， and the abilities to generate images and verify the authenticity of images were strengthened according to the characteristics learned from the dataset respectively. Thirdly， the arbiter was generated by the generator and the discriminator after the last round of adversarial training and metric score calculation module， and the adversarial training results of the generator and the discriminator were measured by this arbiter and fed back into the training plan. Finally， a wining limit was added to the network structure to improve the stability of model training， and the Circle loss function was used to replace the BCE loss function， which made the model optimization process more flexible and the convergence state more clear. Experimental results show that the proposed algorithm has a good generation effect on the architectural and face datasets. On the Large-scale Scene UNderstanding （LSUN） dataset， the proposed algorithm has the Fréchet Inception Distance （FID） index decreased by 1.04% compared with the DCGAN original algorithm； on the CelebA dataset， the proposed algorithm has the Inception Score （IS） index increased by 4.53% compared with the DCGAN original algorithm. The images generated by the proposed algorithm have better diversity and higher quality.

Key words: deep learning, Generative Adversarial Network (GAN), arbitration mechanism, Circle loss, Convolution Neural Network (CNN)

摘要：

针对深度卷积生成对抗网络（DCGAN）中的对抗训练缺乏灵活性以及DCGAN所使用的二分类交叉熵损失（BCE loss）函数存在优化不灵活、收敛状态不明确的问题，提出了一种基于仲裁机制的生成对抗网络（GAN）改进算法，即在DCGAN的基础上引入了所提出的仲裁机制。首先，所提改进算法的网络结构由生成器、鉴别器和仲裁器组成；然后，生成器与鉴别器会根据训练规划进行对抗训练，并根据从数据集中感知学习到的特征分别强化生成图像以及辨别图像真伪的能力；其次，由上一轮经过对抗训练的生成器和鉴别器与度量分数计算模块一起组成仲裁器，该仲裁器将度量生成器与鉴别器对抗训练的结果，并反馈到训练规划中；最后，在网络结构中添加获胜限制以提高模型训练的稳定性，并使用Circle loss函数替换BCE loss函数，使得模型优化过程更灵活、收敛状态更明确。实验结果表明，所提算法在建筑类以及人脸数据集上有较好的生成效果，在LSUN数据集上，该算法的FID指标相较于DCGAN原始算法下降了1.04%；在CelebA数据集上，该算法的IS指标相较于DCGAN原始算法提高了4.53%。所提算法生成的图像具有更好的多样性以及更高的质量。

关键词: 深度学习, 生成对抗网络, 仲裁机制, Circle loss, 卷积神经网络