Multi-image face alignment method and device based on generative adversarial network

Abstract

The invention discloses a multi-image face alignment method based on the generative adversarial network. The method comprises that multiple real face images are input to a generator, and the generatorprocesses the images to a synthesis image which fits distribution of the real images; the real face images and the synthesis image are input to a discriminator to obtain the real probability of the real image and the real probability of the synthesis image, and parameters of the generator and the discriminator are updated iteratively till the parameters are converged so as to determine a model established by the generator and the discriminator; and the face images to be aligned to input to the determined model, and an aligned face image is obtained by once forward transmission operation. Thus, one clear and aligned face image can be generated according to multiple face images, the sparse degree of a noise matrix obtained by subtracting the aligned image from a multi-face-image matrix reflects the degree in which the generated face image is aligned with the input face images, and the generative adversarial network records integral and detailed characteristics of the face.

Description

technical field [0001] The present invention relates to the technical field of multi-image face alignment, in particular to a method and device for multi-image face alignment based on a generative confrontation network. Background technique [0002] Multi-image face alignment technology, that is, to process multiple input face images to produce a technology that makes all output face images have as consistent lighting, occlusion, posture and other conditions as possible. Face alignment technology has a wide range of applications in video surveillance, media processing, public security investigation and other fields. For example, in the face recognition system, by aligning the face images, the face images have the same posture and other conditions, thereby improving the accuracy of face recognition. [0003] The existing multi-image face alignment technologies mainly fall into the following categories: [0004] 1) Face alignment method based on multi-image similarity [1]. ...

AI Technical Summary

Problems solved by technology

In terms of alignment effect, the traditional alignment method can only deal with face alignment of two-dimensional transformation (such as rotation, translation and scaling), and cannot cope with extreme lighting and occlusion conditions and face alignment of three-dimensional pose transformation.

Although the alignment method based on two-dimensional feature points solves the shortcomings of traditional face alignment in terms of efficiency, the face feature points obtained through sample learning are themselves biased estimates of the real feature points of the face.

Especially under extreme conditions, the accuracy of feature point estimation will be greatly reduced, which leads to poor face alignment.

The method based on the 3D face model requires a large number of input images of the same person in different poses, otherwise it is difficult to reconstruct an accurate 3D model, which will also affect the effect of 3D face alignment

Images