|
改进FSRNet的人脸超分辨率网络 |
Face Super Resolution Network with Improved FSRNet |
投稿时间:2020-11-02 修订日期:2020-11-02 |
DOI: |
中文关键词: 人脸超分辨率 注意力损失 先验信息 生成对抗网络 神经网络 |
英文关键词: face super-resolution attention loss a priori information generative adversarial network neural network |
基金项目:国家自然科学基金(61203172);四川省科技计划项目(20ZDYF0008) |
|
摘要点击次数: 279 |
全文下载次数: 0 |
中文摘要: |
人脸超分辨网络FSRNet使用人脸几何先验信息优化人脸超分辨率,可以从低分辨率人脸图像生成逼真的高分辨率人脸图像。但是FSRNet生成的超分辨率图像存在伪影,模糊等不足。我们基于FSRNet,提出了改进的人脸超分辨率网络。我们对FSRNet关键模块进行改进,并引入新的损失函数,以实现更强大的人脸超分辨率网络。直接输入16x16低分辨率图像,最后使用转置卷积放大图像,降低了计算复杂度,并提升了粗略SR网络的性能。通过两步训练,先单独训练粗略SR网络,再训练剩余网络部分,解决网络训练时调参困难的问题。引入热图损失,面部注意力损失和对抗性损失训练,提高了超分辨率人脸图像的质量。实验结果证明,我们改进的方法可以生成面部细节更加清晰的高质量人脸图像。 |
英文摘要: |
Face super resolution network FSRNet use geometric prior information to optimize the super-resolution of human face, which can generate realistic high-resolution face images from low-resolution face images. However, the super-resolution image generated by FSRNet has some shortcomings such as artifact and blur. Based on FSRNet, we propose an improved face super-resolution network. We improve the key module of FSRNet and introduce new loss function to realize a more powerful face super-resolution network. Directly input the low-resolution image (16x16), and finally use Deconv to enlarge the image. The computational complexity is reduced and the performance of coarse SR network is improved. Through two-step training, the coarse SR network is first trained separately, and then the rest of the network is trained, so as to solve the problem of parameter adjustment during network training. Heatmap loss, face attention loss and adversarial loss training are introduced to improve the quality of super-resolution face image. Experimental results show that our improved method can generate high-quality face images with clearer facial details. |
View Fulltext
查看/发表评论 下载PDF阅读器 |
关闭 |
|
|
|