【作者】 杨晓；

【导师】 张文栋； 王红亮；

【作者基本信息】 中北大学 ， 精密仪器及机械， 2009， 硕士

【摘要】 随着计算机技术的高速发展,红外图像目标识别技术得到了广泛关注和应用。利用红外CCD相机对目标进行图像采集,并对采集的图像进行识别是红外图像目标识别的基本任务。红外图像目标识别中最为关键的技术是预处理、特征量提取和分类器设计。红外图像的预处理是红外图像目标识别的第一步,主要包括图像滤波,图像增强,阈值分割。本文使用综合滤波法进行了滤波处理,使用拉普拉斯变换和Butterworth高通滤波进行了图像增强处理,使用迭代法和最小类间方差法对图像进行了阈值分割处理,均取得了令人满意的效果。红外图像特征提取的任务是获得可分性强、稳定性高的数学特征。本文从红外CCD的成像特性、目标的几何特征、图像矩、奇异值变换等角度分别讨论了红外目标图像特征提取的方法。首先,根据目标的几何特征量的要求,通过设计目标的最小外接矩形和选择合适的边缘提取算法等途径实现了对目标几何特征量的提取;进而,利用图像矩提取红外目标的区域形状的特征;再者,使用矩阵的SVD变换提取了目标的尺度奇异值特征,并针对目前存在的问题进行了部分的改进;最后,通过相关的实验,验证了这三种特征提取方法的有效性。分类识别是红外图像目标识别的最后环节,设计性能优越的分类器及其算法是分类能否成功实现的关键。本文主要对最小距离分类器和基于最小错误率的Bayes分类器作了探讨和研究,两种分类方法都取得了良好的分类效果。本文对预处理,特征提取和分类器设计中的各个算法进行了仿真,并验证了这些算法的可行性。更多还原

【Abstract】 As the rapid development of computer technology, infrared image pattern recognition technology based on image processing gets more attentions and broad applications. Target recognition based on infrared CCD comera capturing imagines of the targets is a basic task of the image target recognition. Image preprocessing, feature extraction and classifiers design are most important technologies of image target recognition.Preprocessing is the first step of image target recognition, which has three parties: noise reduction, image enhancement and image segment. In this thesis, the compositive filter method is applied to image filter, Laplace transformation and Butterworth high-pass method are applied to the image enhancement, and Otsu method and Iterative method are applied to the image segment. The preprocess results of these methods are effective.The main mission of feature extraction is obtaining the mathematic features with high divisibility and stability. On the basis of the characteristics of infrared CCD comera, of the target’s shape, of the imagine quadrature and of the SVD transformation, these feature extraction methods are discussed. Firstly, geometry parameter fearures are extracted by means of designing the minimal rect of the target and choicing fit algorithm of edge detection. Secondly, several taeget’s scope shape features are extracted according to the second-order moment. Thirdly, SVD transformation features from target are extracted and some improvement is done according to the existing problem. Finally, these three feature extracting methods are proved effective through relevant experiments.Classification-recognition is final step of image target recognition, so the prominent classifier and algorithms are key steps for classification. In this thesis, the least distance method and Bayes based on least of the rates about mistakes are designed, which are proved effective for targets’classification. All methods of preprocessing, of feature extraction and of classier mentioned in this thesis are simulated, and were proved feasibily.更多还原