基于改进生成对抗网络的动漫人物头像生成算法

Anime Character Portrait Generation Algorithm Based on Improved Generative Adversarial Networks

针对已有的动漫人物头像生成方法中生成结果的多样性较差,且难以准确地按照用户想法按类生成或按局部细节生成的问题,基于含辅助分类器的对抗生成网络(ACGAN),结合互信息理论、多尺度判别等提出了一种改进模型LMV-ACGAN(Latent label attached Multi scale ACGAN with improved VGG mode),用于动漫人物头像的生成。文中设计的模型主要包括特征整合的反卷积生成器,多尺度特征提取器以及真假、类别、隐参数,还原3个全连接神经网络。对于网络结构,所提模型除了类别标签外,额外引入了一组连续值的隐参数,用来增强对模型的约束,同时将卷积神经网络部分的VGG模型中的池化层替换为跨步卷积,并且判别器引入了图像的多尺度信息进行特征融合且改进了网络末端结构以及各部分的参数更新方式,以尽可能减弱末端的分类部分、真假判别部分和隐参数还原部分之间的相互影响。实验结果表明,所提模型有效地解决了模式崩塌的问题,同时较ACGAN提高了模型生成指定类型图像的成功率和准确度,对于ACGAN等生成失败或者类型判别错误的图像,可以做到正确生成,且能够通过调整连续的隐参数有效地实现一些简单的图像编辑功能,如人脸的朝向等。

In order to solve the problems of poor diversity,generation by class and detail control in existed method,we present an improved model named LMV-ACGAN.It is based on ACGAN and involved with mutual information and multiscale discrimination.Our model includes a feature combined generator,a multiscale discriminator and three fully connected nets for real-fake judging,classifying and latent label restoration.As a semi-supervised generative model,except class label,we also use a group of continuous latent label to enhance the constraint of the generator.Moreover,in our algorithms,pooling layers in VGG-NET are replaced by stride convolutions.Then the discriminator uses the multiscale information of the image to feature fusion.Finally,we improve the tail-end structure of the model and the rules of parameters update so as to reduce the influence between classification,real-fake judgement and latent label restoration as far as possible.Our experiment shows that the proposed method effectively solve the problem of mode collapse on our dataset,meanwhile compared with origin ACGAN,our method increases the success rate and accuracy of generating specified class image.For the image which is generated poorly or classified incorrectly by ACGAN,our method can achieve the goal.In addition,our model enable people to modify the continuous latent label to realize image editing such as changing the face orientation.