解耦表征学习视角下认知图像属性特征的图像生成方法  

Image Generation Method for Cognizing Image Attribute Features from the Perspective of Disentangled Representation Learning

作　　者：蔡江海 黄成泉[1,2,3] 王顺霞 罗森艳 杨贵燕 周丽华 CAI Jianghai;HUANG Chengquan;WANG Shunxia;LUO Senyan;YANG Guiyan;ZHOU Lihua(Key Laboratory of Pattern Recognition and Intelligent Systems of Guizhou Province,Guizhou Minzu University,Guiyang 550025;School of Data Sciences and Information Engineering,Guizhou Minzu University,Guiyang 550025;Engineering Training Center,Guizhou Minzu University,Guiyang 550025)

机构地区：[1]贵州民族大学贵州省模式识别与智能系统重点实验室,贵阳550025 [2]贵州民族大学数据科学与信息工程学院,贵阳550025 [3]贵州民族大学工程技术人才实践训练中心,贵阳550025

出　　处：《模式识别与人工智能》2024年第7期638-651,共14页Pattern Recognition and Artificial Intelligence

基　　金：国家自然科学基金项目(No.62062024);贵州省科技计划项目(No.黔科合基础-ZK[2021]一般342);贵州省研究生教育教学改革重点项目(No.黔教合YJSJGKT[2021]018);贵州省教育厅自然科学研究项目(No.黔教技[2022]015);贵州省模式识别与智能系统重点实验室2022年度开放课题(No.GZMUKL[2022]KF03)资助。

摘　　要：在生成式人工智能领域,解耦表征学习的研究进一步推动图像生成方法的发展,但现有的解耦方法更多地关注图像生成的低维表示,忽略目标变化图像内在的可解释因素,导致生成的图像容易受到其它不相关属性特征的影响.为此,文中提出解耦表征学习视角下认知图像属性特征的图像生成方法.首先,从生成模型的潜在空间出发,通过训练获得关于目标变化图像的候选遍历方向.然后,构建无监督语义分解策略,并基于候选遍历的方向联合发现嵌入在潜在空间中的可解释方向.最后,利用解耦编码器和对比学习构建对比模拟器和变化空间,进而由可解释方向提取目标变化图像的解耦表征并生成图像.在5个解耦数据集上的实验表明文中方法性能较优.In the field of generative artificial intelligence,the research of disentangled representation learning further promotes the development of image generation methods.However,existing disentanglement methods pay more attention to low-dimensional representation of image generation,ignoring inherent interpretable factors of the target variation image.This oversight results in generated image being susceptible to the influence of other irrelevant attribute features.To address this issue,an image generation method for cognizing image attribute features from the perspective of disentangled representation learning is proposed.Firstly,candidate traversal directions for the target variation image are obtained by training,starting from the latent space of the generative model.Secondly,an unsupervised semantic decomposition strategy is constructed,and the interpretable directions embedded in the latent space are jointly discovered based on the direction of candidate traversals.Finally,a contrast simulator and a variation space are constructed using disentangled encoders and contrastive learning.Consequently,the disentangled representations of the target variation image are extracted from the interpretable directions and the image is generated.Extensive experiments on five popular disentanglement datasets demonstrate the superior performance of the proposed method.

关 键 词：解耦表征学习 潜在空间 可解释方向 图像生成 变化空间 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]