Earlier this year, the Chinese multinational technology and entertainment conglomerate company, Tencent, was reported to have applied for a patent for a training method, device, equipment, and storage medium for a facial forgery model. Now, the company has been granted a new patent for facial counterfeiting, which can be applied to assisted driving.
The patent specification states that as artificial intelligence technology advances, its study and use in the field of face counterfeiting also grow. Most deep learning-based face identification models now rely on labelled images of fictitious faces for supervised training.
However, based on a small number of pseudo-face photos, only a few pseud-face images have label information. The face identification model developed by image training does not have a high level of identification accuracy.
The current application, which relates to the technical area of artificial intelligence, provides a training procedure, apparatus, tool, and storage medium for a face forgery model.
Elegant Themes - The most popular WordPress theme in the world and the ultimate WordPress Page Builder. Get a 30-day money-back guarantee. Get it for Free
The process entails: obtaining a fake face image and a genuine face image and ensuring that there is no difference more than or equal to a threshold value between the facial poses of the two images; The pseudo-face image and the real face image are fused to create a fused face image using the gradient information associated with the pseudo-face image. The face recognition model is then trained using the identification results associated with the fusion of facial images.
At the same time, it is clear from the patent abstract that the embodiments of the present application can be used in situations involving assisted driving, intelligent transportation, and artificial intelligence. Based on the combination of facial photos and label data, the model’s self-supervised learning is accomplished without being constrained by the lack of samples, increasing the model’s forgery detection precision.