一种飞机图像目标多特征信息融合识别方法  

An Airplane Image Target's Multi-feature Fusion Recognition Method

作　　者：李新德[1] 杨伟东[1] DEZERT Jean 

机构地区：[1]东南大学自动化学院复杂工程测量与控制教育部重点实验室 [2]法国国家航天航空研究中心

出　　处：《自动化学报》2012年第8期1298-1307,共10页Acta Automatica Sinica

基　　金：国家自然科学基金(60804063;61175091);江苏省自然科学基金(BK2010403);图像信息处理与智能控制教育部重点实验室开放基金(200902);东南大学创新基金(3208000501);东南大学优秀青年教师教学;科研资助计划(3208001203)资助;航空科学基金(20100169001)~~

摘　　要：提出了一种基于概率神经网络(Probabilistic neural networks,PNN)和DSmT推理(Dezert-Smarandache theory)的飞机图像目标多特征融合识别算法.针对提取的多个图像特征量,利用数据融合的思想对来自图像目标各个特征量提供的信息进行融合处理.首先,对图像进行二值化预处理,并提取Hu矩、归一化转动惯量、仿射不变矩、轮廓离散化参数和奇异值特征5个特征量;其次,针对DSmT理论中信度赋值构造困难的问题,利用PNN网络,构造目标识别率矩阵,通过目标识别率矩阵对证据源进行信度赋值;然后,用DSmT组合规则在决策级层进行融合,从而完成对飞机目标的识别;最后,在目标图像小畸变情形下,将本文提出的图像多特征信息融合方法和单一特征方法进行了对比测试实验,结果表明本文方法在同等条件下正确识别率得到了很大提高,同时达到实时性要求,而且具有有效拒判能力和目标图像尺寸不敏感性.即使在大畸变情况下,识别率也能达到89.3%.This paper proposes an image target＇s multi-feature fusion recognition method based on probabilistic neural networks （PNN） and Dezert-Smarandache theory （DSmT）. To aim at multiple features extracted from an image, the information from them is fused. Firstly, the image is preprocessed with binarization and then multiple features are extracted, such as Hu moments, normalized moment of inertia, affine invariant moments, discrete outline parameters and singular values. Secondly, due to the difficulty of the construction of the basic belief assignment in DSmT, in this paper the target recognition rate matrix is constructed by PNN, that is, the basic belief assignments can be assigned to the evidence sources by PNN. Finally, the procedure of airplane target recognition can be accomplished by the DSmT combination rule at the level of decision fusion. For small distortion of target image, the multi-feature fusion method proposed in this paper is compared with the single-feature one through a series of experiments. The experimental result in this paper proves that this method greatly improves the right recognition rate, satisfies real-time requirements, and has good ability of rejection of judgement and strong insensitivity to target image size. And even for big distortion, the right recognition rate can also reach 89.3%.

关 键 词：信息融合 目标识别 Dezert-Smarandache推理 飞机图像 概率神经网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]