面向语义特征的距离损失函数研究  

The research of semantic feature-oriented distance loss function

作　　者：林益文 陈青[1] 邱新媛 王俊 雷鹏英 LIN Yi-wen;CHEN Qing;QIU Xin-yuan;WANG Jun;LEI Peng-ying(Nuclear Power Institute of China,Chengdu 610213,China)

机构地区：[1]中国核动力研究设计院,成都610213

出　　处：《信息技术》2024年第12期141-148,共8页Information Technology

基　　金：中国核工业集团公司集中研发项目(ZNZBA02)。

摘　　要：卷积神经网络在图像识别过程中具有相当的可解释性,能够以一种可理解的方式对相关特征概念进行提取。文中探索一种利用卷积神经网络可解释性的方法,并提出一种面向语义特征的距离损失函数,迫使基础模型学习富有差异的目标特征概念,同时使用模型融合方法对相关语义特征进行整合以识别目标对象。实验以流行的Cifar10数据集和VGG16网络为基准,从初始化、距离函数、分割阈值、融合方法和数据增强等关键组成对方法框架进行了摸索,明确了相关组成部分对方法框架的影响机制,证实了该方法对图像识别能力具有良好的提升效果。Convolutional Neural Network have considerable interpretability in the image recognition process,enabling the extraction of relevant feature concepts in an understandable way.This paper explores an approach to exploit the interpretability of Convolutional Neural Network,and proposes a semantic feature-oriented distance loss function.It forces the base models to learn different target feature concepts and the model fusion approach integrates the relevant semantic features to recognize the target object.The current Cifar10 data set and VGG16 network are used as benchmarks,the method framework is explored from key components such as initialization,distance function,segmentation threshold,fusion method and image enhancement.As a result,the influence mechanism of relevant components on the method framework is clarified,and the good improvement effect of the method on image recognition ability is confirmed.

关 键 词：卷积神经网络 语义特征 距离函数 可解释性 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]