基于改进的生成对抗网络的动漫头像生成算法

Animation Head Sculpture Generation Algorithm Based on Improved Generative Adversarial Networks

针对大部分生成对抗网络在动漫图像的生成上会呈现出训练不稳定,生成样本多样性比较差,人物局部细节上效果不好,生成样本质量不高的问题,文章利用条件熵构造的一种距离惩罚生成器的目标函数,结合注意力机制提出一种改进模型MGAN-ED。模型主要包括融入多尺度注意力特征提取单元的生成器和多尺度判别器。采用GAM和FID进行评估,所做实验结果表明模型有效地解决了模式崩塌的问题,生成图像的局部细节更加清晰,生成样本质量更高。

In view of the problems of training instability,poor diversity of generated samples,poor effect on local details of characters and low quality of samples generated in most of the Generative Adversarial Networks on generation of the animation head sculptures,this paper constructs a distance penalty generator target function by using conditional entropy,and an improved model MGAN-ED is proposed combined with Attention Mechanism.The model mainly includes a generator integrated with multi-scale attention feature extraction unit and a multi-scale discriminator.The GAM and FID are used to evaluate the model.The experimental results show that the model can effectively solve the problem of pattern collapse,and the local details of the generated image are clearer and the quality of the generated samples is higher.