Face super-resolution method combined with 3D face structure prior

Abstract

The invention discloses a face super-resolution method combined with 3D face structure priori. The method can explicitly combine with 3D face priori to capture high-definition face structure information and provide 3D topological information based on face attributes for a network, such as identity, expression, texture, brightness and face posture. The invention provides a deep learning network framework, and the framework generally comprises two branches; an upper half branch comprises a ResNet-50 network for mining human face 3D information from an input image, and combining and reconstructing the human face 3D information into a face rendering structure; and a lower half part branch utilizes a spatial domain feature conversion layer, combines 3D information and a face rendering structure as 3D prior, and utilizes a space attention mechanism and a channel attention mechanism to realize face super-resolution; The prior can be embedded into any network, the performance is effectively improved, and convergence is accelerated.

Description

technical field [0001] The invention belongs to the field of computer vision human face super-resolution, and in particular relates to a human face super-resolution method combined with 3D facial structure prior. Background technique [0002] Face image information has many applications in computer analysis in today's society, such as face recognition and medical diagnosis. But often various technologies require face images to have a higher resolution. When the resolution of the face image is relatively low, the technical accuracy drops dramatically. Therefore, the face super-resolution algorithm came into being to help restore face images from low resolution to high resolution. [0003] Today's advanced face super-resolution algorithms usually use deep convolutional networks to learn the mapping relationship between low-resolution and high-resolution face patterns. However, most methods do not make full use of facial structure and identity information and are difficult t...

AI Technical Summary

Problems solved by technology

Therefore, face super-resolution, especially at high magnification, remains a challenging problem

Images