【作者】 金野；

【导师】 唐降龙；

【作者基本信息】 哈尔滨工业大学 ， 计算机科学与技术， 2008， 博士

【摘要】 纸币图像分析理解技术对纸币流通监管具有十分重要的意义。本文对多光谱纸币图像进行深入分析,研究了快速纸币分类、纸币特征检测(污损特征、防伪特征)、纸币质量评估,设计完成了纸币图像分析系统,并应用到不同的实际产品中。在快速纸币分类中,采用分步识别策略。首先采用“网格-GMM”方法进行预分类,然后采用“类Haar-AdaBoost”分类器区分相似纸币。在预分类过程中,提取网格特征,然后采用基于结构风险最小化(SRM)的混合高斯模型(GMM)构造分类器,以提高识别的稳定性。在相似纸币区分中,首先提出了类Haar特征的提取方法,然后采用AdaBoost算法选取有效特征,以提高分类器的区分能力。实验证明,该分步策略显著提高了系统对于低质量纸币的处理能力,大幅降低了系统拒识率。在纸币特征检测中,采用基于图像配准/匹配的检测算法。在第一种基于边缘特征的算法中,首先采用基于图像区域的配准算法,以确定待检图像与参考图像之间的对应关系。然后采用Kirsch算子计算两图像的边缘强度差(EID),并根据EID来提取缺损特征。该算法对于图像上的新增边缘十分敏感,而对整体亮度偏移则较为稳定,在处理折旧纸币时,拥有较好的鲁棒性。在第二种算法中,提出首先采用均匀性特征确定污损待检象素,然后采用基于均匀性特征的图像配准算法确定待检图像与参考图像间的象素对应关系,最后将污损待检象素与其参考象素进行比较,以确定缺损状况并进行质量分析。在第三种方法中,将均匀性特征推广到彩色空间,取得了更为精确的质量分析结果。在纸币退化分析中,首次详细讨论了纸币的自身特点,并据此对纸币退化及形变(D&D)模型进行了研究。将纸币形变采用基于B-样条的FFD网格来表示,将纸币退化采用颜色混合模型(CDM)来表示。在CDM中,纸币退化被看作“全局磨损”与“局部缺损”共同作用的结果,并据此将参考图像到待检图像的颜色偏移分解为磨损折旧及缺损位移,分别进行评估。在第一种算法中,构建纸币退化能量函数(BDE)作为数据驱动项,进行基于FFD模型的图像配准,实现对纸币退化的量化分析。第二种算法中,对于每一类纸币,将原有的静态参考图像扩展为动态的退化图像序列,将纸币质量评估看作在相应序列中寻找对应帧的过程。利用3维FFD模型,同时考虑纸币的退化及形变,显著地提高了纸币分析的整体性能。在纸币分析系统构建中,首先介绍了高速多光谱接触式图像传感器(CIS)的设计原理,通过该传感器可以依次采集纸币的红、绿以及红外光谱图像,然后提出了“FPGA+DSP+上位机”的纸币分析系统构架。在硬件配置中,讨论了DSP的外设配置以及FPGA的逻辑设计,包括触发控制,曝光控制,行采集以及I/O端口设计等。在软件设计中,首先讨论了图像采集流程,然后提出了传感器补偿算法,用于补偿CIS各采集单元的成像误差,最后讨论了图像分析流程。在图像分析流程中,在DSP片内实现实时性的基本功能,如纸币分类,缺损/防伪特征检测以及质量评估,将其他的扩展功能留在上位机实现。采用该构架设计了两款纸币分析系统,第一款用于CF3000型纸币清分机,该系统中,通过数个图像分析模块的组合,实现对纸币的全面分析。第二款用于WJD_TKTH07型纸币点钞检伪仪,该系统中,将数项功能集成于同一处理模块。实验证明,该构架设计灵活,可适用于各种不同的应用场合。更多还原

【Abstract】 Banknote image analysis is of vital importance to the currency supervision. In this thesis, the muti-spectrum images of banknote are systematically investigated for the first time. The algorithms of banknote image analysis are first studied including banknote classification, feature detection/verification (defect feature or anti-counterfeit feature), and quality evaluation, based on which, the system of banknote image processing is then achieved and applied to different enverinments.In banknote classification, a hierarchical recognition strategy is proposed. First, the banknotes are pre-classified by the Grid-GMM Method, and then the Haar-Adaboost classifiers are used for the identification between the similar banknotes. During the pre-calssification, Grid Features are adopted, and a Gaussian mixture model (GMM) based on structural risk minimization (SRM) is engaged to build a more robust classifer. During the similar banknote identification, a kind of Haar-like feature is proposed, which is then refined by the Adaboost method to gain high performance on the classification of the similar images. The exeperimental result reveals that this hierachical method leads to a high capacity for low quality banknote processing and greatly decreases the false reject rate.In banknote feature detection/verification, three algorithms based on image registration/match are proposed. In the first edge-based algorithm, an area-based image registration algorithm is adopted to overlay the sensed and reference banknote images, in which an Edge Intensity Differential (EID) of the two images is constructed from the edge information extracted by the Kirsch operator. The Defect Feature extracted by EID is sensitive to the odd edge-information, and is robust to the global intensity change, which makes it suit for the attrited banknote. In the second algorithm, the homogeneity feature of banknote are introduced to locate the pixels that probably been blurred, the image registration algorithm based on the homogeneity feature is subsequently used to overlay the sensed and reference paper currency image. At last, each probably polluted pixel on the sensed image is compared with its corresponding pixel on the reference image to estimate the deterioration level. In the third algorithm, the homogeneity feature is extended to the chromatic space, which gains a more accurate result on banknote quality evaluation.In banknote deterioroation analysis, the characteristics of banknotes are discussed in detail for the first time. Based on it, a deformation and deterioration (D&D) model of banknote are proposed, in which the banknote deformation is interpreted by a cubic B-spline based Free Form Deformation (FFD) grids and the banknote deterioration is described by a Color Diffusion Model(CDM). In the CDM, banknote deterioration is regarded as a joint effect of General Attrition and Local Defect, and accordingly, the color shift from the sensed image to the reference image can be decomposed into the Attrition Rate and Defect Distance and evaluated separately. Based on the D&D model, two banknote evaluation algorithms are proposed. In the first one, a Banknote Deterioration Energy (BDE) is proposed as the data-driven term in FFD-based image registration, by which the quantitative analysis of banknote deterioration achieved. In the second one, for each kind of banknote image, the static reference image is extended to a dynamic image sequence that changed according to the deterioration degree. Then, the quality evaluation of the sensed image can be interpreted as locating its optimal position in the corresponding sequence. In this three-dimension FFD model, the deterioration degree and deformation state of one banknote can be concerned simultaneity, which will improve the banknote analysis performance greatly.In the banknote analysis system, a high speed Muti-spectrum Contact Image Sensor (CIS) is first introduced, by which one banknote can be alternately sampled under the red, green, and infrared lights, and then a“FPGA+DSP+PC”framework for the banknote analysis system is proposed. In the hardware configuration, the peripherals of the DSP and the logic setting of the FPGA are discussed, including trigger control, exposure contral, line sampling, and the I/O ports setting. In the software design, the image sampling flow are first discussed, and then a sensor compensation algorithm is proposed to equalize the sensitive error of each sampling unit in CIS, and at last, the image analysis scheme is discussed. In the image analysis scheme, the real-time basic function is implemented inside the DSP, such as banknote classification, defect/anti- counterfeit feature detection and quality evaluation, and the other extended function can be implemented in PC. Using this framework, two banknote analysis systems are constructed. The first one is designed for the CF3000 banknote sorter, in which several image analysis modules are associated to achieve a full analysis of the banknotes. The second one is designed for WJD_TKTH07 anti-counterfeit banknote counter, in which several functions are integrated into one image analysis module. The experinmental result reveals that this framework is flexible enough to be applied to different conditions.更多还原