Multi-angle side face frontage method based on feature mapping

Abstract

The invention discloses a multi-angle side face frontage method based on feature mapping. According to the method, firstly, a face recognition model Light CNN is used as a feature extractor, deep features of a side face input image and a real front face image are extracted, a mapping relation between side face features and front face features is learned, and a model M is obtained; and secondly, a generative adversarial network (GAN) is used as a main network, a side face image is used as an input image, a model M is used for mapping side face features into front face features, meanwhile, an encoder in the GAN is used for extracting the side face image features, the two features are spliced in the channel dimension to serve as input features of a decoder, and finally a vivid virtual front face image is output. According to the method, the identity information retention of the generated image is improved on the face global feature and the eye circumference local feature, the method can be used as a preprocessing process to help to improve the performance of a face recognition model, and especially, the recognition rate is improved on the multi-angle side face under the monitoring view angle.

Description

technical field [0001] The invention relates to a multi-angle side face frontalization method based on feature mapping, and belongs to the field of side face frontalization methods. Background technique [0002] With the continuous development of artificial intelligence, face recognition has been widely used in security systems, video surveillance and other fields. Faces in the surveillance scene will have multiple posture changes, such as frontal faces, side faces with deflection angles, and side faces with changes in both deflection and depression angles. Due to the self-occlusion problem of the side face, some key identity features in the face are missing, and it is difficult to directly use the side face and the front face for recognition. In order to alleviate this problem, the frontalization method is generally adopted, that is, using a single or multiple side face images to restore the virtual frontal face and then recognize it. This method can be used as a preproce...

AI Technical Summary

Problems solved by technology

[0004] The method based on the 3D model requires mutual mapping between the 2D image and the 3D model. This process will introduce some errors, and the frontalization results are usually not realistic enough, and there will be artificial artifacts and serious loss of facial texture information.

Although the result recovered by the two-dimensional method is relatively smooth, the identity information of the obtained frontal image is relatively low

In addition, the existing methods mainly focus on the frontalization of the side face in the deflection direction, and pay less attention to the side face that includes both deflection and depression angle changes under the monitoring perspective.

Images