基于主成分分析的多传感器目标识别技术研究

发布时间：2018-06-22 14:37

本文选题：目标识别 + LBP　；参考：《沈阳理工大学》2017年硕士论文

【摘要】：目标识别是一个综合了图像处理、机器学习和模式识别等领域的课题,是近几年的一个研究热点。本文主要是对基于主成分分析的多传感器目标识别技术进行研究,为了提高多传感器的目标识别率,提出了一种基于LBP-PCA的多传感器目标识别算法。主要研究内容如下:首先,本文对目标识别中的图像预处理以及图像描述进行了研究。图像预处理相关方法中重点研究了图像滤波、增强以及分割等相关环节。在图像滤波及增强环节中本文选择的是中值滤波和直方图均值化算法,并用实验检验了这两种算法对本文采集到的图像的适用性;图像分割环节主要包括一阶微分算子以及阈值分割算法,本文对多种分割算法进行分析和研究,通过实验效果的比较选取Otsu阈值分割作为本文的分割算法。图像描述部分主要研究了各种特征提取算法,并分析了各种算法的优缺点,其中主要包括了图像的颜色、形状、纹理等特征提取方法。其中,对本文研究的LBP纹理特征进行了简要说明,具体研究过程放在后面的章节。对本文后续对比实验用到的SIFT描述子进行研究,证明其适用于本文提出的算法,能够有效的进行特征描述。其次,本文选取支持向量机用于决策判断。比较了几种分类器的优缺点。主要研究了支持向量机的分类原理,包括线性可分支持向量机、线性不可分支持向量机、非线性支持向量机及其核函数的选择和支持向量机的应用与拓展。再次,对基于主成分分析的异类传感器融合算法进行研究。分析了三种融合方式,对三种融合方式的优缺点进行比较并结合本文算法的要求,选取出特征级信息融合方式。在传感器类型的选择上,结合实际情况以及对各种传感器的适用范围选取红外传感器和可见光传感器作为本文研究对象。重点对提出的主成分分析融合算法的各个方面进行研究,主要包括PCA的定义、基本原理以及推导过程。并且通过与传统串联法以及单一传感器的目标识别实验进行对比,可以验证出PCA算法可以有效地降低特征向量的维度,减少运算量的同时保留原始数据的主要信息。最后,本文提出了一种基于LBP-PCA的多传感器目标识别算法,并进行了实验验证,首先将红外以及可见光图像分别通过相应的预处理算法从图像中提取出目标,然后提取目标的LBP特征点向量,再利用PCA算法降低提取出的LBP特征向量的维数从而得出融合后的特征向量,最后利用SVM进行识别与分类用以获取识别率。实验仿真结果表明LBP特征提取方法具有良好的旋转不变性和灰度不变性,同时主成分分析可以从一个高维空间中的提取主要的特征,利用LBP-PCA多传感器目标识别算法可以克服传统图像融合中数据过大、运行时间过长等问题,在实现实时检测的同时提高了目标的识别率。
[Abstract]:Target recognition is a research hotspot in recent years, which integrates the fields of image processing, machine learning and pattern recognition. In order to improve the rate of multi-sensor target recognition, a multi-sensor target recognition algorithm based on LBP-PCA is proposed in this paper. The main contents are as follows: firstly, the image preprocessing and image description in target recognition are studied in this paper. In the correlation method of image preprocessing, image filtering, image enhancement and image segmentation are studied. The median filter and histogram mean algorithm are selected in image filtering and enhancement. The applicability of these two algorithms to the images collected in this paper is verified by experiments. Image segmentation mainly includes first order differential operator and threshold segmentation algorithm. This paper analyzes and studies many segmentation algorithms. Otsu threshold segmentation is selected as the segmentation algorithm by comparing the experimental results. In the part of image description, various feature extraction algorithms are studied, and their advantages and disadvantages are analyzed, including color, shape, texture and other feature extraction methods. Among them, the LBP texture features studied in this paper are briefly described, and the specific research process is put in the following chapter. The SIFT descriptors used in the subsequent comparative experiments in this paper are studied, and it is proved that the proposed algorithm can effectively describe the features. Secondly, support vector machine is selected for decision-making. The advantages and disadvantages of several classifiers are compared. This paper mainly studies the classification principle of support vector machine, including linear separable support vector machine, linear non-separable support vector machine, nonlinear support vector machine and its kernel function, and the application and development of support vector machine. Thirdly, the fusion algorithm of heterogeneous sensors based on principal component analysis (PCA) is studied. The advantages and disadvantages of the three fusion methods are analyzed and the feature level information fusion method is selected according to the requirements of this algorithm. In the selection of sensor type, the infrared sensor and the visible light sensor are selected as the research object of this paper, combined with the actual situation and the applicable range of various sensors. All aspects of the proposed principal component analysis (PCA) fusion algorithm are studied, including the definition, the basic principle and the derivation process of PCA. By comparing with the traditional series method and the target recognition experiment of a single sensor, it can be verified that PCA algorithm can effectively reduce the dimension of the eigenvector, reduce the computation amount and retain the main information of the original data. Finally, a multi-sensor target recognition algorithm based on LBP-PCA is proposed and verified by experiments. Firstly, infrared and visible images are extracted from the image by corresponding preprocessing algorithms. Then the LBP feature point vector of the target is extracted, then the dimension of the extracted LBP feature vector is reduced by PCA algorithm, and the fused feature vector is obtained. Finally, SVM is used for recognition and classification to obtain the recognition rate. Experimental results show that LBP feature extraction method has good rotation invariance and gray invariance, and principal component analysis can extract the main features from a high-dimensional space. The LBP-PCA multi-sensor target recognition algorithm can overcome the problems of too large data and long running time in traditional image fusion. It can achieve real-time detection and improve the target recognition rate.
【学位授予单位】：沈阳理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP212

【参考文献】