PSWGAN-GP:改进梯度惩罚的生成对抗网络

PSWGAN-GP:Improved Wasserstein Generative Adversarial Network with Gradient Penalty

生成对抗网络的出现对解决深度学习领域样本数据不足的研究起到了极大的促进作用。为解决生成对抗网络生成的图像出现轮廓模糊、前景背景分离等细节质量问题,提出一种改进梯度惩罚的Wasserstein生成对抗网络算法(PSWGAN-GP)。该算法在WGAN-GP的Wasserstein距离损失和梯度惩罚的基础上,在判别器中使用从VGG-16网络的3个池化层中提取的特征,并通过这些特征计算得出风格损失(Style-loss)和感知损失(Perceptual-loss)作为原损失的惩罚项,提升判别器对深层特征的获取和判别能力,对生成图像的细节进行修正和提升。实验结果表明,在生成器和判别器网络结构相同,并保证超参数相同的情况下,PSWGAN-GP的IS评分和FID评分相对于参与对比的其他图像生成算法有所提升,且可有效改善生成图片的细节质量。

The emergence of generative adversarial network(GAN)plays a great role in solving the problem of insufficient sample data in the field of deep learning.In order to solve the detail quality problems of images generated by GAN such as foreground and background separation and contour blurring,this paper proposes an improved Wasserstein generative adversarial network with gradient penalty(PSWGAN-GP)method.Based on the Wasserstein distance loss and gradient penalty of WGAN-GP,this method uses the features extracted from the three pooling layers of the VGG-16 network in the discriminator and calculates the style-loss and perceptual-loss from these features as penalty terms of the original loss,which improves the discriminator’s ability to acquire and discriminate deep features and enhance the details of the generated images.The experimental results show that PSWGAN-GP can effectively improve the quality of generated images with the same generator and discriminator network structure and the same hyperparameters,and the scores in IS and FID are improved relative to other image generation methods.