基于生成对抗网络的人脸线条画自动生成
首发时间:2020-09-04
摘要:人脸线条画旨在用简单线条表达人脸的关键特征,具有极高的艺术价值。当前自动生成人脸线条画的研究还处于探索阶段,本文首次提出基于生成对抗网络的人脸线条画生成算法。为克服线条画中左右脸特征的较大差异,并减少背景对线条生成的影响,本文在生成器中引入了自注意力机制,以捕获全局的"面部特征依赖性"。为得到更清晰的线条画,降低噪声的影响,本文提出了一个长循环结构,该模块建立了从假人脸到线条画的映射,并通过循环一致损失为原始生成器提供反馈信息。在CHUK-LineDraw公开数据集上的测试结果显示:与改进前的方法及同类方法相比,本文方法取得了更优的视觉效果,获得了最低的FID得分;与真实线条画相比,SIFT距离得分也最低。
关键词: 生成对抗网络 循环一致性生成对抗网络 人脸线条画 注意力机制
For information in English, please click here
Automatic Generation of Face Line Drawings based on Generative Adversarial Network
Abstract:Face line drawings aim to express the key features of a human face with simple lines, which has extremely high artistic value. The current research on automatically generating face line drawings is still in the exploratory stage. For the first time, this paper proposes a face line drawing generation algorithm based on a generative confrontation network. In order to overcome the large difference between the left and right facial features in the line drawing, and reduce the influence of the background on the line generation, this paper introduces a self-attention mechanism in the generator to capture the global "facial feature dependence". In order to obtain a clearer line drawing and reduce the impact of noise, this paper proposes a long loop structure, which establishes a mapping from the fake face to the line drawing, and provides feedback information for the original generator through the loop consistency loss. In addition, the application of dual training strategy makes full use of paired data and unpaired data, alleviating the problem of missing data. The test results on the CHUK-LineDraw public dataset show that: Compared with the method before the improvement and similar methods, the method in this paper has achieved better visual effects and obtained the lowest FID score; compared with the real line drawing, the SIFT distance is also the best.
Keywords: generative adversarial networks cycle-consistent generative adversarial network face line drawing attention mechanism
基金:
引用
No.****
同行评议
勘误表
基于生成对抗网络的人脸线条画自动生成
评论
全部评论0/1000