梯度控制与鉴别特征引导的写实类人物肖像插画转换方法

Gradient Controlled and Discriminative Features Guided Image-to-Image Translation Method Towards Realistic Portrait Illustrations

现有的无监督图像转换方法由于未考虑人脸辨别特征保持这一问题,转换后得到的写实类人物肖像插画常会出现人脸变形和面部结构坍塌的现象,难以辨认人物信息.针对该问题,文中提出梯度控制与鉴别特征引导的写实类人物肖像插画转换方法.在循环生成对抗网络(CycleGAN)的基础上引入避免冗余特征复用的掩码残差长连接,将图像梯度信息一致性作为约束条件,较好地保持人脸辨别特征.设计鉴别特征引导的信息共享训练机制,使生成器具有和鉴别器相同的提取目标风格图像鉴别特征的能力.同时拓展图像块鉴别器为多感知鉴别器,获得丰富的鉴别信息.实验表明,文中方法转换得到的写实类人物肖像插画不仅较好地保持显著的人脸辨别特征,而且在插画视觉效果上较优.

Without considering the preservation of facial recognition characteristics,the existing unsupervised image-to-image translation methods often suffer from face distortion and facial structure collapse.Consequently,it is difficult to identify the personal information after translation.To tackle this issue,gradient controlled and discriminative features guided image-to-image translation method towards realistic portrait illustrations is proposed based on cycle-consistent generative adversarial network(CycleGAN).Masked residual long connections are introduced by the proposed method to avoid reusing redundant features and image gradient information consistency is considered as a constraint to preserve facial recognition characteristics.In addition,a discriminative feature guided information-shared training mechanism is devised and thus generators are capable of capturing discriminative features of target images,like the discriminators.Moreover,patch-level discriminators are extended to multi-awareness discriminators to obtain more discriminative information.Experimental results show that the proposed method preserves facial recognition characteristics well in the translated illustrations and outperforms the existing unsupervised image-to-image translation methods in visual effect of illustrations.