Face image normalization method based on auto-encoding network

Abstract

The invention provides a face image normalization method based on an auto-encoding network, which comprises the following steps: step 1, constructing a training data set of the auto-encoding network,and preprocessing the training data set; 2, constructing a coding network and a decoding network of the self-coding network, wherein the coding network is constructed based on a Resnet34 module, and the decoding network is composed of a deconvolution module; 3, using an L1loss loss function to measure the difference between the noise face and the normalized face, using a cross entropy loss function to measure the face image classification loss, and using the result of weighted summation of the two losses as the final loss function of the self-encoding network; step 4, training the preprocessedtraining data set on a self-encoding network to obtain a trained face normalization model; and step 5, inputting the to-be-processed face image into the face normalization model to complete face image normalization. By the adoption of the technical scheme, faces can be normalized in batches at a time without any prior information, compared with other face image restoration algorithms, the methodhas the advantages of being simple, convenient, easy to implement, high in efficiency and good in effect, face image identity information before and after normalization is kept consistent, has the advantages of being invariant in feature and high in accuracy and is of great significance to a face recognition algorithm.

Description

technical field [0001] The invention relates to the field of image processing, in particular to a face image normalization method based on a self-encoding network. Background technique [0002] The massive face image data in real life contains a lot of interference information unrelated to the face, which will affect the accuracy of the face recognition algorithm and hinder the judgment of face identity information. These interference information mainly include facial occlusions, decorations, image damage, exaggerated expressions of characters, profile, and light. Existing algorithms have separately studied these interfering factors that make face recognition difficult, such as algorithms for removing occluders and repairing damaged images, algorithms for correcting exaggerated expressions of characters, algorithms for removing light interference on images, and algorithms for correcting side effects. face algorithm. Since these algorithms can only remove a certain item of ...

AI Technical Summary

Problems solved by technology

These algorithm models are complex, have single functions, and remove less types of interference information. At present, there is no unified framework that can remove all kinds of interference at the same time and realize end-to-end face normalization.

In a

Images