Face super-resolution reconstruction method and system based on semantic features

Abstract

The invention discloses a face super-resolution reconstruction method and system based on semantic features. The method comprises the following steps: step 1, a quality reduction stage: synthesizing high-resolution faces of a training set into a low-resolution face image containing complex noise and fuzzy distribution; and step 2, a generation stage: forming a super-resolution face reconstruction result as the output of the network, the system comprising a synthesis module, an amplification module and an integration module, the generalization ability of the FSR network model in the multi-degradation mode is improved, and the SR face perception quality reconstructed in the multi-degradation mode is enhanced at the same time.

Description

technical field [0001] The present invention relates to a face super-resolution reconstruction method and system based on semantic features. Background technique [0002] Super-resolution reconstruction (SR) is an important area of ​​image quality enhancement research. This technique can recover more texture information from low-resolution (LR) images and form high-resolution (HR) images. Face super-resolution reconstruction (FSR) is an applied branch of SR technology. This branch is dedicated to recovering high-resolution face shapes from low-resolution faces. FSR technology can be applied to assist related advanced vision tasks, such as face recognition, face correction, and the application of related biological features such as security video analysis. [0003] Deep learning techniques have been widely explored in face super-resolution reconstruction. Based on the research on the application of convolutional neural network structure in natural image SR reconstruction,...

AI Technical Summary

Problems solved by technology

[0005] Under the existing technology, supervised methods require paired LR-HR face datasets to support model training, and LR faces are obtained from real faces through a fixed downsampling method. This single degradation process assumes that it cannot Adapt to face reconstruction requirements in multiple degraded modes

In unsupervised learning, without the reference of a real face, the reconstructed SR face will be deformed or distorted, which will affect the visual effect.

At the same time, in the multi-degradation mode, some prior information such as facial landmarks and edges is inaccurate or unobtainable, and the reconstruction effect of the FSR model guided by this is not good.

Images