【作者】 徐华；

【导师】 马社祥；

【作者基本信息】 天津理工大学 ， 信号与信息处理， 2009， 硕士

【摘要】 随着计算机技术的不断发展,伪造印章印鉴的难度大大下降,从而导致伪造印鉴引起的犯罪活动屡禁不止。印鉴识别算法的研究,其目标是以高技术水平进行科学的、准确的文件鉴定,向社会提供一个具有自动化的高技术保障。本论文主要针对印鉴识别算法进行了研究。论文中首先分析了印鉴识别的难点和重点,然后针对这些难点和重点,提出了一种新的印鉴配准方法并提取了四种互补的特征。这种新的印鉴配准算法首先求出印鉴的几何中心,以几何中心为圆心,先在极径方向上抽样离散化,然后离散化每一个同心圆,重采样每个同心圆上的点,最后以小区域匹配求出旋转角度,完成印鉴的配准。实验表明,这种方法是有效的。针对印鉴的特点,我们提取了四种不同的,互补的印鉴特征和各自的分类方法。这四种特征分别是:基于区域的形状特征提取、基于印鉴配准的特征提取、基于小波变换的纹理特征提取和改进的SIFT特征提取。基于区域的形状特征提取是利用Hu矩不变量提取印鉴的形状,然后对不同形状的印鉴进行分类;基于印鉴配准的特征提取是在印鉴配准的基础上,定义一个差异性函数,根据这个函数选取一个阈值判断印鉴的真伪;基于小波变换的纹理特征提取,主要是针对印鉴中的细纹理特征,依据低频子带的能量判断小波变换的级数,对小波变换后的子带图像提取均值和方差,作为特征向量输入支持向量机进行印鉴真伪的鉴别; SIFT是一种稳定的,匹配能力强的算法,针对SIFT实时性差,本论文提出了基于改进的SIFT特征提取,改进的SIFT是在印鉴配准的基础上,将标准印鉴和待测印鉴按角度等划分,使特征向量的搜索范围变小,从而减少匹配的时间。另外,根据印鉴识别的特点,印鉴识别取决于边缘点,背景点和非边缘点对印鉴的影响不大,我们将印鉴配准后再提取边缘,然后利用SIFT算法计算特征点,利用欧式距离分类匹配点。实验表明,改进的SIFT算法在匹配率和实时性上都要优于SIFT。最后,本文介绍了多分类器融合的基础知识,采用投票规则将这四种不同的印鉴识别方法进行了融合。实验结果证明这四种特征是有效的,互补的。更多还原

【Abstract】 With the development of the computer techniques, difficulties of forging seal decline. Therefore economic crimes caused by forgery seals are hard to be forbidden. The purpose of seal identification study is to provide a system that can identify tested seal images to be genuine or forgery, and it also can offer an automatic, efficient and accurate documentation authority identification way to avoid crimes.This paper focuses on Seal identification. First, discuss the emphasis and difficulties of seal identification, Based on these emphasis and difficulties, a new registration method and four complementary features are used. This new registration method fit all kinds of seal. It first get the center of mass, then use the center of mass as center, make the radius and concentric circle discretization and resample the points in concentric circle, At last get the angle’s value using small region matching and finish the registration. This method is proved efficient by experiment.According to the feature of seal, this paper brings up four different and complementary features and their different classification methods. These features are shape based on Hu Moment Invariants, difference based on seal registration, texture based on wavelet and ISIFT. Shape features are used to classify the seal of different shape by abstracting Hu Moment. We define a difference function based on seal registration. According the difference function, we choose a threshold value to realize seal identification. Texture features based on wavelet choose mean and variance of all sub-channels as an input into SVM. SIFT is a stable and powerfully matching method. Because SIFT costs so much time, this paper gives an improved SIFT method to decrease the time cost. ISIFT divides the seal into several regions after registration and edge extracting. The small region makes searching cost much less time. In addition, the points on edge are key factor to determine seal identification, we can ignore other points. Then use Euclidean distance to judge feature points after registration and edge extracting. The experiment tells ISIFT is better than SIFT in aspect of rate of registration and real-time.更多还原