Low-resolution face image super-resolution method for recognition.

Abstract

The invention discloses a low-resolution face image super-resolution method for recognition. The method comprises the following steps: learning a degradation distribution rule of sample data by adopting a variational self-encoding network through a small number of collected monitoring scene degradation samples; on the basis of the data degradation distribution rule, expanding diverse degradation samples obeying the data degradation distribution rule through a sampling strategy; performing sample degradation simulation generation processing on the high-resolution data sample through a style migration network so as to obtain a corresponding low-resolution face sample which is distributed in the same way as the real degradation sample; and training a preset face image super-resolution networkby using the high-resolution data sample and the correspondingly generated low-resolution degraded sample until the network converges. According to the invention, the face super-resolution enhancement network not only can generate high-quality clear faces, but also pays attention to synthesis of identity information related textures. According to the method, the recognition rate of the low-resolution face image in a monitoring scene can be improved under the condition that the number of low-resolution samples is small.

Description

technical field [0001] The invention relates to the technical field of face image super-resolution, in particular to a face recognition method based on face super-resolution preprocessing. Background technique [0002] Face recognition technology has the advantages of non-mandatory, non-contact and concurrent processing. In practical application scenarios, it can meet the filtering, authentication and recognition of multiple faces at the same time. It has been widely used in buildings, communities, shopping malls and subways. Video surveillance and financial payment security certification of mobile clients such as mobile phones and other life fields. [0003] In recent years, with the development of deep learning, the recognition accuracy of face recognition technology has been qualitatively improved, and its recognition accuracy has surpassed the accuracy of human naked eye recognition on the authentication data set LFW. However, these mainstream face recognition algorithm...

AI Technical Summary

Problems solved by technology

Its main purpose is to solve the problem that the current face image super-resolution method only focuses on image visual quality improvement, but does not explicitly focus on identity information recovery, and improves the recognition accuracy of super-resolution face images

Images