【作者】 盖杉；

【导师】 唐降龙；

【作者基本信息】 哈尔滨工业大学 ， 人工智能与信息处理， 2011， 博士

【摘要】 纸币是一个国家的名片,流通纸币的清洁程度反映了一个国家的文明程度,体现了一个国家的实力与地位。纸币图像包含了可见光图像信息、红外光图像信息、紫外光图像信息以及磁图像信息、纸张厚度信息等多类别多模态信息。通过对纸币图像的分析与理解能够高可靠地对纸币进行分类,并且检测出假币、旧币、残币以及不易流通的纸币,保证流通纸币的安全性、可靠性和整洁性。本文重点研究了纸币图像分析的几个关键技术,并构建了一个实际纸币清分系统,具体研究内容如下:(1)首先针对传统纸币图像特征提取方法中掩模特征稳定性欠佳与网格特征不易区分风格相近纸币图像的缺陷,研究了基于离散Haar小波变换与模糊逻辑相结合的纸币图像特征提取方法。首先使用Haar小波对纸币图像进行分解操作,提取出纸币图像的低频小波系数与高频小波系数。在此基础上通过引入模糊逻辑的方法来描述纸币图像的灰度模糊性,分别把提取的小波系数作为语言变量,输入到相应的隶属度函数中,在模糊特征空间中求出每个模糊区域对应的激活强度值,将这些激活强度值进行归一化处理后构成纸币特征向量,最后使用神经网络分类器对纸币图像进行识别。提取的纸币特征向量具有较好的可区分性与抗干扰性,解决了低质量纸币图像如污损图像、受到噪声干扰的图像以及扭曲变形纸币图像的识别不一致性问题。(2)针对小波变换不具备灵活的方向选择性,而且不能最优地稀疏表示图像的缺陷,提出了一种基于Contourlet变换与模糊逻辑相结合的纸币图像特征提取方法。由于纸币图像具有丰富的纹理结构特征信息,因此通过将纸币图像分解为不同尺度不同方向的子带来达到区分不同纸币图像纹理特征区域的目的。解决了低质量纸币图像识别性能较差的问题,具有高稳定性与可靠性。根据纸币图像经过Contourlet变换后各子带系数分布的统计特性,研究了基于Contourlet变换与混合高斯模型相结合的纸币图像分类方法,该方法首先对纸币图像进行Contourlet变换,采用混合高斯模型描述纸币图像变换系数的统计分布;然后运用最大似然估计算法训练模型参数,并且将训练得到的模型参数集合作为特征向量进行识别。该方法首次将统计建模思想引入到纸币识别过程中。(3)为了能更好地捕获纸币图像纹理特征信息与局部相位特征信息,研究了基于旋转四元数小波变换的纸币图像分类方法。旋转四元数小波变换是通过重新构造四元数小波变换中的Garbor滤波器来实现的。其中旋转四元数小波变换由一个幅值与三个四元相位组成,其中两个四元相位表示纸币图像的局部位移信息,第三个四元相位表示纸币图像纹理信息,幅值表示四元相位信息变化趋势,同时四元数旋转小波变换具有平移不变性。首先运用旋转四元数小波变换对纸币图像进行分解操作;然后计算每个分解子带系数的能量与标准差;最后运用支持向量机进行纸币图像识别。该方法取得了较高的识别率并且能够满足清分系统的实时性要求。(4)为了提高纸币图像污损检测的准确率,同时减少撕裂与笔迹等现象对污损检测造成的影响,提出了一种基于小波变换的纸币图像污损检测方法。首先采用基于小波变换的图像配准等价性框架对纸币图像进行配准;然后运用Kirsch边缘检测算子提取纸币图像的边缘信息,将计算得到的纸币图像边缘幅值差作为污损特征;最后将纸币图像划分为若干固定大小子区域,通过对每个子区域的污损特征统计来判断该区域是否存在污损。提取的污损特征对于纸币图像灰度值退化现象具有较强的抗干扰能力,同时具有高污损识别率与检测稳定性。(5)在上述研究内容基础上,本文完成了一个实际的纸币清分系统并且已经投入到实际应用中。更多还原

【Abstract】 Banknote is a country business card, the clean degree of banknote in circulation reflects civilization degree and can embody the strength and status of country. The banknote image includes multi-class and multi-mode information such as visible image information, infrared image information, ultraviolet image information, magnetic image information and paper thickness information. Banknote classification and detection of counterfeit currency, defected currency, worn currency and unease circulation currency are completed by analysis and understanding of banknote high reliably. Doing that in order to make circulation safty, reliability and netness. Several key techniques of banknote image analysis are studyied in this paper, and construct a practical banknote sorting system. The really research content is as follows:1. According to the defects of feature extraction methods which has low stability of mask and difficulty discrimination of grid feature, a new banknote feature extraction method based on Haar wavelet transform and fuzzy logic is proposed. Firstly, apply the Haar wavelet transform to the banknote image, and then obtain the approximation and detail coefficients. The fuzzy of banknote image is discribed by fuzzy logic. We make two linguistic variables corresponding two coefficients, and the firing strength is calculated by membership function in the fuzzy feature space. Then the banknote image feature vector is obatined by normalizing the firing strength. Finally, the neural network is applied to classify the banknote image. The extracted feature has sensitiveness and robustness. It is well solved the classification inconformity caused by defected image, noise and distoration during the sample by contact image sensor.2. The wavelet transorm has two drawbacks which are non sensitive direction selection and non sparse reprensentation of banknote image. So the new banknote feature extraction method based on Contourlet transform and fuzzy logic is proposed. The rich textual information of banknote image is extracted by decomposing the banknote image into different directions and resolutions. It has good recognition ability to low quality banknote image. Meanwhile the feature extraction method is posed based on statistical characteristics of coefficients which are at different resolution and direction. The Contourlet transoform is used to decompose the banknote image, and then using the mixture gaussian model describes the coefficients distribution, using the EM algorithm to estimate the parameters of the model. The idea of statistical modeling is applied to banknote image classification firstly.3. In order to capture the rich textual information and local shift information of banknote image, the new banknote image classification method based on rotated quaternion wavelet transform is proposed. The new rotated quaternion wavelet transform is constructed by changing the Garbor filter. The rotated quaternion wavelet consists of one magnitude and three phases, the two phase represent local shift information of image ant the other denote the textual information of image. Firstly, apply the rotated quaternion wavelet transform to the banknote image; then calculate the standard deviation and energy. Finally, using the support vector machine to classify the banknote image. The method obtains high recognition rate and satisfy the real-time requirement of banknote sorting system.4. In order to improve the defect detection rate and decrease the effects of tearing and handwriting, a new banknote image defect detection method is proposed based on wavelet transform. The affine transformation and wavelet transformaton are used to image registration. The edge detector is applied to extract the edge information of banknote image. The banknote image defected feature is obtained by calculating the difference of magnitude. Then the whole banknote image is divided into several same rectangular region, and judge the defect banknote image via regional defect feature. The proposed method has strong resistence to gray scale alteration, and obtain high recognition rate and stability.5. A practical banknote sorting system is finished based on research contents described above and make it to real application.更多还原