基于类混合高斯映射的归纳式广义零样本识别  

INDUCTIVE GENERALIZED ZERO-SHOT LEARNING BASED ON GAUSSIAN-MIXTURE-LIKE MAPPING

作　　者：白万荣 张驯 张蕾 杨凡 邵洁 Bai Wanrong;Zhang Xun;Zhang Lei;Yang Fan;Shao Jie(Gansu Electric Power Research Institute,Lanzhou 730050,Gansu,China;College of Electronics and Information Engineering,Shanghai University of Electric Power,Shanghai 200090,China)

机构地区：[1]国网甘肃省电力公司电力科学研究院,甘肃兰州730050 [2]上海电力大学电子与信息工程学院,上海200090

出　　处：《计算机应用与软件》2024年第11期206-212,共7页Computer Applications and Software

基　　金：国家自然科学基金项目(61802250);国网甘肃省电力公司2019科技开发项目。

摘　　要：在广义零样本识别研究中,分类器对可见类别的偏倚以及在高维向低维特征映射过程中产生的信息丢失是传统算法常见的两大问题。为了解决上述问题,基于高斯混合分布模型的思想,结合共同学习的设计理念,提出一种加权多通道结构,不仅能够通过建立通道学习速率差异化实现通道间有监督的共同学习,而且可以通过类多高斯分布的计算拟合生成特征的真实分布特性,增强网络在隐藏空间的特征映射能力。针对三个常用数据库CUB、AWA2和SUN进行测试后,实验结果表明,基于多通道和生成对抗网络建立的类混合高斯映射网络模型针对上述三个数据库调和指标H分别提高了1.4、1.56和0.47,验证了这种加权多通道结构实现的类混合高斯映射模型在广义零样本图像识别领域的有效性。In the study of generalized zero-shot learning,the partial dependence of the classifier on the visible class and the information loss in the process of high-dimensional to low-dimensional feature mapping are two main problems in traditional algorithms.In order to solve the problems,based on the idea of Gaussian mixture distribution model and combined with the design concept of common learning,this paper proposes a multi-channel structure.The structure can not only realize supervised common learning between channels by establishing channel learning rate differentiation,but also can fit the real distribution characteristics of the generated features through the calculation of multi-Gaussian distribution and enhance the feature mapping capability of the network in hidden space.In order to verify the multi-channel structure,this paper conducted a large number of experiments on three benchmark databases CUB,AWA2 and SUN.Harmonic index H has increased 1.4,1.56 and 0.47 for I-GZSL.It proves the effectiveness of the multi-channel structure in generalized zero shot learning.

关 键 词：广义零样本识别 归纳式 高斯混合映射 多通道结构 图像识别 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]