MeGAN:基于多任务增强生成对抗网络的图像合成

MeGAN:Multi-task enhanced generative adversarialnetwork for image synthesis

在计算机视觉领域,现有图像合成方法通常采用一对一的映射网络生成人脸表情,存在很大的建模局限性,难以表达丰富多样、复杂多变的人脸表情。为此,该文提出一种基于多任务增强生成对抗网络的图像合成方法。该方法构建多任务学习框架,改善人脸表情生成的多样性;通过设计双域卷积模块,利用具有补偿的频域信息改善空域特征映射;引入多尺度自适应激活函数,对不同特征进行自适应修正,进一步提升网络性能和特征映射效果。实验结果表明,该文方法能够同时生成多种逼真的人脸表情图像,与现有先进的图像合成方法相比,具有更好的定性和定量评估结果。

In computer vision fields,existing image synthesis methods usually feature a one-to-one mapping network to generate facial expressions.But they have the inherent limitations,which hinder the accurate representation of diverse facial expressions.For this reason,a novel multi-task enhanced generative adversarial network(MeGAN)for facial image synthesis is proposed.This network adopts a multi-task learning framework to improve the diversity of facial expression generation.The dual-domain convolution module is designed to use frequency-domain features as complementary information for improving the learning of spatial feature mapping.A multi-scale adaptive activation function is introduced to modify the feature maps adaptively for further improvement of network performance.Experimental results show that the proposed method can generate a variety of realistic facial expression images simultaneously and usually achieve better qualitative and quantitative results than the state-of-the-art methods.