提高生成样本的质量和多样性一直是生成对抗网络(generative adversarial network, GAN)领域主要挑战任务之一. 为此, 提出了一种双向约束生成对抗网络(bidirectional constraint generative adversarial network, BCGAN). 与传统GAN变体相比, 该网络在架构设计上增加了一个生成器模块, 两个生成器分别从两个不同方向逼近真实样本的数据分布. 然后根据BCGAN的网络架构, 设计了新的损失函数, 并对其进行了理论分析及证明. 在BCGAN的训练过程中, 一方面通过增加两个生成样本数据分布之间的距离来丰富生成样本的多样性, 另一方面通过减小鉴别器对两个生成样本数据分布之间的差异来稳定训练过程, 提高生成样本的质量. 最后, 在1个合成数据集和3个不同公开挑战数据集上进行了实验. 一系列实验证明, 较其他生成方法相比, 所提方法对真实数据分布具有更强的拟合能力, 能够有效提升生成样本的质量和多样性. 此外, 所提方法的训练过程更加平滑稳定.
Improving the quality and diversity of generated samples has always been one of the main challenging tasks in the field of generative adversarial network (GAN). For this reason, a bidirectional constraint GAN (BCGAN) is proposed. Compared with the traditional GAN variants, this network adds one more generator module to the architecture design. The two generators approach the data distribution of real samples from two different directions. Then, according to the network architecture of BCGAN, this study designs a new loss function and analyzes and proves it theoretically. During BCGAN training, the diversity of the generated samples is enriched by increasing the distance between the data distribution of two generated samples, and the difference of the discriminator between the data distribution of the two generated samples is reduced to stabilize the training process and thereby improve the quality of the generated samples. Finally, experiments are carried out on a synthetic dataset and three open challenge datasets. This series of experiments show that compared with other generative methods, the proposed method fits real data distribution better and effectively improves the quality and diversity of generated samples. In addition, the training process of this method is smoother and more stable.