CN-115908608-B - Face personalized paper-cut generating method based on constraint cycle generation countermeasure network

CN115908608BCN 115908608 BCN115908608 BCN 115908608BCN-115908608-B

Abstract

The application discloses a face personalized paper-cut generating method based on constraint cycle generation countermeasure network, which comprises the steps of S10, establishing a face data set comprising a face image and a paper-cut data set comprising a paper-cut image, S20, obtaining a corresponding face analysis data set according to the face data set by adopting a pre-training model, distinguishing a key area and a non-key area by black and white colors for the face analysis image in the face analysis data set, S30, designing a cycle countermeasure network for characteristic fusion of the face key area, S40, training the cycle generation countermeasure network by adopting the face data set, the paper-cut data set and the face analysis data set, obtaining a face generation paper-cut generator model for fusing the face key area after the cycle generation countermeasure network fused with the key area constraint is stable, and S50, inputting the face image and the face analysis image into the face generation paper-cut generator model to obtain the face personalized paper-cut.

Inventors

ZHOU HOUPAN
ZHANG TAIXIANG
HUANG JINGZHOU

Assignees

杭州电子科技大学

Dates

Publication Date: 20260505
Application Date: 20221111

Claims (3)

1. The face personalized paper cut generation method based on constraint circulation generation countermeasure network is characterized by comprising the following steps: S10, establishing a face data set comprising a face image and a paper-cut data set comprising a paper-cut image; S20, obtaining a corresponding face analysis data set according to the face data set by adopting a pre-training model, and distinguishing a key area and a non-key area by using black and white colors in a face analysis image in the face analysis data set; S30, designing a cyclic countermeasure network for face key region feature fusion, wherein the cyclic countermeasure network for face key region feature fusion adopts two-stage fusion of face key regions, and the face key region feature fusion assists in generating paper-cut in the face generation paper-cut stage; The design process of the cyclic countermeasure network for the feature fusion of the key areas of the human face comprises the following steps: Firstly, a generator G_A for generating a paper cut from a human face is established, and a generator G_B for generating the human face from the paper cut is generated, wherein G_A consists of G_A_conv and G_A_tran, G_A_conv represents a convolution layer of the paper cut generator generated from the human face, G_A_tran represents a deconvolution layer of the paper cut generator generated from the human face, G_B consists of G_B_conv and G_B_tran, G_B_conv represents a convolution layer of the human face generator generated from the paper cut, G_B_tran represents a deconvolution layer of the human face generator generated from the paper cut, G_A_conv is added with the convolution layer Conv_3 to obtain convolution data G_A_conv_data of a human face image, and G_B_conv is added with the convolution layer Conv_3 to obtain convolution data G_B_conv_data of the paper cut image; Secondly, extracting high-level semantic features in the face image by adopting a feature model to obtain convolution data E_A_data, and respectively carrying out feature fusion on the G_A_conv_data and the G_B_conv_data and the E_A_data to obtain new convolution data G_A_conv_new_data and G_B_conv_new_data; Thirdly, the convolution data G_A_conv_new_data and G_B_conv_new_data are respectively input into G_A_tran and G_B_tran to obtain a paper-cut image and a face image; then obtaining a cyclic countermeasure network fused with the key area characteristics of the face; S40, training the cyclic generation countermeasure network by adopting a face data set, a paper-cut data set and a face analysis data set to obtain a face generation paper-cut generator model fusing key areas of the face; And S50, inputting the face image and the face analysis image into the face generation paper-cut generator model to obtain the personalized paper-cut of the face.
2. The method of generating personalized paper cuts for faces based on constrained loop generation countermeasure network of claim 1, wherein the key areas include hair, eyes, nose, and mouth.
3. The method for generating personalized paper cuts on a human face based on constrained loop generation countermeasure network according to claim 1, wherein P-Net is used as a pre-training model.

Description

Face personalized paper-cut generating method based on constraint cycle generation countermeasure network Technical Field The application relates to the field of computer vision image conversion, in particular to a face personalized paper cut generation method based on constrained circulation generation of a countermeasure network. Background The paper-cut is art created by using paper as a processing object and scissors as a tool, is a first national non-material cultural heritage, and the personalized paper-cut of a human face is one of the paper-cuts. Under the background of the current national advanced traditional culture, the creation potential of the personalized paper-cut of the face is infinite, but the non-intelligent method is unfavorable for large-scale creation, and the inheritance and popularization of the personalized paper-cut of the face are limited, so that the intelligent generation of the personalized paper-cut is a new way for inheriting and developing the non-genetic culture. With the rapid development of deep learning in recent years, it has become possible to generate personalized paper cuts using image-to-image conversion techniques. At present, three methods exist for personalized paper-cut generation of human faces. First, processing is performed by computer aided design tools, such as using photoshop software. Secondly, calibrating and extracting facial features of the face of the person to be cut by using a face cutting template mode, for example, ASM (active shape model, calibration and separation component for a target face), and generating a target face cutting chart by using a deformation method. Thirdly, generating paper-cut which is similar to the facial features by generating a countermeasure network mode and utilizing a generator and a generator after the countermeasure learning of a discriminator is optimized. The method has poor intelligence, time and labor waste, and excessive noise is treated by adopting a generated countermeasures network, so that the obtained facial paper-cut has large difference with the facial features of the real face, and the personalized requirement cannot be met. Disclosure of Invention The application aims to provide a face personalized paper-cut generating method based on a constraint circulation generation countermeasure network, which generates paper-cut which is similar to the facial sense of the human face based on a face key area after face image analysis. The face personalized paper-cut generating method based on the constrained circulation generation countermeasure network comprises the following steps of establishing a face data set comprising face images and a paper-cut data set comprising paper-cut images, S20, obtaining a corresponding face analysis data set according to the face data set by adopting a pre-training model, distinguishing key areas and non-key areas by black and white colors for the face analysis images in the face analysis data set, S30, designing the circulation countermeasure network with characteristic fusion of the face key areas, S40, training the circulation generation countermeasure network by adopting the face data set, the paper-cut data set and the face analysis data set to obtain a face generation paper-cut generator model fusing the face key areas, and S50, inputting the face images and the face analysis images into the face generation paper-cut generator model to obtain the face personalized paper-cut. Preferably, the critical area includes hair, eyes, nose and mouth. Preferably, the cyclic countermeasure network for the feature fusion of the key areas of the human faces adopts two-stage fusion of the key areas of the human faces, the feature fusion of the key areas of the human faces assists in the generation of paper cutting in the stage of generating paper cutting of the human faces, and the feature fusion of the key areas of the human faces assists in the generation of the human faces in the stage of generating the paper cutting. The method for generating the personalized facial cut based on the constraint cycle to generate the countermeasure network according to the claim 3 comprises the steps of firstly establishing a generator G_A for generating the paper cut from the facial surface, and generating a generator G_B for generating the facial surface from the paper cut, wherein the G_A consists of a G_A_conv and a G_A_tran, the G_A_conv represents a convolution layer of the paper cut generator generated from the facial surface, the G_A_tran represents a deconvolution layer of the paper cut generator generated from the facial surface, the G_B consists of a G_B_conv and a G_B_tran, the G_B_tran represents a deconvolution layer of the facial surface generator generated from the paper cut, the G_A_conv is added with convolution data G_A_conv of a face image, the G_A_conv is obtained by the convolution layer, the G_A_conv is obtained by the convolution data of the G_A_conv, the G_A_conv is obtained by