Improved GAN Model for Image Animation
Sirong Re, Jiaxin Li, Yiran Li, Junan Mao · 2022 IEEE 5th International Conference on Information Systems and Computer Aided Education (ICISCAE) · 2022
Image-to-image translation is a meaningful and challenging task in computer vision and artistic style transfer. It aims to learn a function which can transfer an image with other style. Animation of real human face photos is one of the most popular image-to-image tasks. Our paper proposed a model which was based on U-GAT-IT for human face image animation task. U-GAT-IT is a GAN-based model which can efficiently transfer the style of picture from different domains. It uses a novel method to implement unsupervised image-to-image translation. Furthermore, detecting face key points is one of the important steps to generate animated face images. Using DLIB is an efficient way to achieve this. However, when faces were occluded by something such as hands and masks, the results were terrible. In our model, a practical facial landmark detector (PFLD) was selected to locate human faces. As PFLD has a special loss function, this model can direct human faces more accurately. Moreover, this change enables to realize the target of image-to-image even when faces were blocked. It could validly distinguish between human face and veils. The experimental results show that our model can obtain good animation effect, and can get high-quality generation results even when the face is occluded.