基于语义关联推理的非合作目标识别  

Semantic association inferencing for non-cooperative object recognition

作　　者：陈璐 石光明 梁宇平 吴金建 CHEN Lu;SHI Guangming;LIANG Yuping;Wu Jinjian(School of Artificial Intelligence,XiDian University,Xi'an 710126,China;Pengcheng Laboratory,Shenzhen 518055,China)

机构地区：[1]西安电子科技大学人工智能学院,西安710126 [2]鹏城实验室,深圳518055

出　　处：《中国体视学与图像分析》2023年第2期134-142,共9页Chinese Journal of Stereology and Image Analysis

基　　金：国家自然科学基金重点项目(No.61836008)。

摘　　要：非合作目标智能识别技术在态势感知中发挥着越来越重要的作用,该类目标通常表现出小样本特性。然而,现有的深度学习方法在受限环境中的数据获取与可解释性方面存在应用瓶颈。与之相比,大脑在先验知识的指导下,快速提取目标语义基元,并借助语义关联完成小样本下的目标精准识别。针对于此,本文构建了深度特征与语义基元间的相互映射关系,并提出了一种基于语义关联推理的非合作目标识别方法。该方法首先定义并提取封闭环境中非合作目标的语义基元、关联客观实景特征和主观虚景语义,并依据先验知识构建知识驱动的语义关联图谱,有效地缓解了模型对于数据的依赖。同时,该方法借鉴大脑知识联想机制,在推理过程中通过对关联图谱中的相关节点进行激活与判定,实现了可解释的非合作目标识别。在非合作目标数据集上的实验证明,本文所提出的方法在精确率和可解释性上具有显著的优越性。The intelligent recognition technology of non-cooperative objects plays an increasingly important role in situational awareness.Typically,non-cooperative objects exhibit few-sample characteristics.Nevertheless,current data-driven deep learning methods are constrained by data acquisition and interpretability issues in limited environments.In contrast,a brain accomplishes few-shot learning through knowledge association:under the guidance of prior knowledge,it quickly extracts object semantic primitives and exploits semantic association to accurately identify objects under few-sample situations.Drawing on the knowledge association mechanism of brains to achieve few-shot learning is a major challenge at present.To address this issue,this paper constructs a mapping relationship between deep features and semantic primitives,and proposes a non-cooperative object recognition method based on semantic association inferencing(SAI-NOR).SAI-NOR effectively alleviates the dependence on data of the model by extracting semantic primitives,associating objective real scene features with subjective virtual scene semantics,and constructing a knowledge driven semantic association graph.Furthermore,SAI-NOR uses the brain knowledge association mechanism for reference,and realizes interpretable non-cooperative object recognition by activating the relevant nodes in the association graph during inferencing.

关 键 词：目标识别 小样本学习 语义关联 知识驱动 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]