作者简介:胡学娟(1981—),女(汉族),湖北省安陆市人,深圳大学助理研究员、博士. E-mail:xjhu@szu.edu.cn
中文责编:方 圆; 英文责编:海 潮
深圳市激光工程重点实验室,深圳大学电子科学与技术学院, 广东省高校先进光学精密制造技术重点实验室,深圳 518060
Hu Xuejuan, Ruan Shuangchen, Guo Chunyu, and Liu ChengxiangShenzhen Key Laboratory of Laser Engineering, College of Electronic Science and Technology, Key Laboratory of Advanced Optical Precision Manufacturing Technology of Guangdong Higher Education Institutes, Shenzhen University, Shenzhen 518060, P.R.China
infrared image; Chinese RMB currency recognition; histogram of oriented gradients; support vector machine; Fisher criterion; precision counterfeit detection; image feature extraction
DOI: 10.3724/SP.J.1249.2014.05487
提出一种基于改进梯度方向直方图和支持向量机分类器的人民币识别方法.利用人民币红外图像中斑马线特征进行真伪识别,通过Fisher准则进行特征块选择实现梯度方向直方图特征的降维.针对斑马线防伪图案进行实验. 结果表明,该方法能克服红外图像中的背景干扰和噪声,得到较好鉴伪结果.
This paper presents a method to improve the histograms of oriented gradient descriptors and support vector machine classifier for Chinese RMB currency recognition. The zebra-stripe pattern of the infrared images of RMB paper currency was used for real and counterfeit classification. The dimension of histograms of oriented gradient features is decreased by feature block selection based on the Fisher criterion. Several experiments on zebra-stripe pattern recognition were conducted, and the proposed method shows its robustness against background interference and noise.
Special inks that absorb infrared(IR)light are used in printing paper currency, which could be an important cue for counterfeit identification. Security ink that absorbs near-IR(NIR)light is an organic functional dye made up of one or several NIR absorption materials. This ink absorbs light waves with wavelengths ranging from 700 nm to 1 100 nm. When used as a localized portion of a printed product, this ink is invisible in daylight and can only be observed with a detection apparatus[1]. The NIR-absorbing material is synthesized under high temperature, which is a highly technical and costly process. Thus, NIR security ink is difficult to forge and effective for anti-counterfeit guarantee.
IR security inks have been widely used by many banknotes. For example, local IR security is used in the dollar, euro, and Portuguese banknotes. By imaging in NIR light, part of the pattern of several different banknotes is hidden selectively. For example, the pattern in Italian banknotes is almost entirely hidden except for the serial number. The IR security pattern of the 2005 edition of the renminbi(RMB)and the Comorian franc not only disappears but also shows another pattern.
Through IR transmission imaging, a zebra-stripe anti-counterfeit pattern invisible in sunlight is presented in the security line area of the 2005 edition of the 10, 20, 50 and 100 RMB notes. Counterfeit paper currency has been reported to show a realistic watermark, an optical variable ink, an invisible denomination, magnetic and ultraviolet characteristics[2], but no forged IR zebra stripes. Thus, the special characteristics of the IR anti-counterfeit method bring unique advantages for paper currency discrimination. Despite that some works have addressed the problem of money recognition, to the best of our knowledge, there are only a few works dealing with the problem of the authentication of the money.
In Ref.[3], light transmittance and an instance-based classifier by the Euclidean distance metric are used to classify the value of banknote. Feature extraction using wavelet Transform and the minimum Euclidean distance matching is described in Ref.[4] for Korean Won bill classification, where image is acquired in the visible light spectra. In Ref.[5-6], Neural network and genetic algorithm have been exploited to address the problem of banknote recognition. In Ref.[7], US banknote recognition has been proposed by sped up robust features. However, these approaches in the above references do not take into account the problem of banknote counterfeit detection, which is the main issue addressed in this paper.
Ref.[8] addresses the problem of Sterling banknote validation through a cascade of segmentation and classification procedures. But the applicability is not straightforward in the context of the RMB, because this currency encodes different strategies to avoid forgeries. In Ref.[9], a method based on multiple-kernel support vector machines is proposed to detect counterfeit Taiwanese currency. Banknotes are firstly divided into partitions, and the luminance of histograms of the partitions is taken as the input of the system. In Ref.[10], the color domain has been considered for US banknote validation. In Ref.[11], the euro banknote image is acquired in the visible light spectrum and two types of neural networks are used for classification. While all of the approaches in Ref.[9-11] work in the visible spectrum, they are actually not robust as most of the tampered banknotes visually look like the genuine one. In Ref.[12], a NIR camera is used to acquire the Euro banknotes images. The average gray value feature and simple thresholding strategy were used for authentication. However, simple average gray value feature descriptors are not robust for geometric and photometric deformation of image. And histograms of oriented gradients(HOG)features are local and robust feature descriptors, which have a disadvantage of higher feature dimensions.
In this paper, the NIR images of RMB currency are acquired for counterfeit detection. To represent the discriminative zebra-stripe anti-counterfeit pattern, improved histograms of oriented gradients are proposed for feature extraction. Support vector machines(SVM)are used for classification. A database with 540 images captured from 500 real notes and 40 counterfeit notes are used for training and testing and as high as 99.03% accuracy has been achieved.
The remainder of the paper is organized as follows: section 1 describes the preprocessing of NIR image of the 100 RMB paper currency and summarizes the proposed algorithms for HOG feature extraction and feature block selection. In section 2, the designed counterfeit detection method is described, whereas experimental results are presented. Finally, conclusions are given.
In this paper, the detected samples are the 2005 edition of 100 RMB paper currencies. An IR image acquisition system as described in Ref.[13] is used to capture the IR transmission image of the paper currencies, so we can distinguish the real money from the counterfeit by the recorded zebra-stripe anti-counterfeit pattern of the IR imaging, including a set of alternating bright and dark rectangular blocks. Fig.1(a)shows an example of the detected pattern, which consists of a group of alternating bright and dark blocks with a cycle of 2×H, 2 times of the width of the bright or dark blocks, and a black line crossing normally the zebra-stripe pattern. In our image preprocessing, if no solid security line appears in the image, the currency is directly recognized as the counterfeit. Otherwise, we locate the horizontal ordinate of the region of interest(ROI)center by the security line, thus two 16-pixel wide rectangle regions were segmented from the left and right sides of the security line to produce the zebra-stripe pattern of interest. Following that, a region with size 5H×32 can be extracted as ROI for feature extraction shown in Fig.1(b).
图1 安全线区域和感兴趣区域的红外图像HOG was introduced by Dalal and Triggs in 2005[14] and has been successfully used in pedestrian detection. The main principle of HOG is that shape characteristics can be properly described by the density distribution of the gradient or edge direction. To extract HOG feature, the ROI region(32×128)extracted was divided into a number of blocks consisting of 2×2 cells, where each cell is composed of 8×8 pixels. For each cell, gradient direction and magnitude were calculated using the gradient operator[-1, 0, 1] shown by
α(x,y)=tan-1[(I(x,y+1)-I(x,y-1))/(I(x+1,y)-I(x-1,y))](1)
and
G(x,y)={[I(x+1,y)-I(x-1,y)]2+
[I(x,y+1)-I(x,y-1)]2}1/2(2)
where I(x,y)is the pixel value at position(x, y)in the image, α(x,y)indicates the gradient direction of the pixel, and G(x,y)indicates the gradient amplitude. An HOG can then be calculated for each cell by weighted voting of every pixel. In this paper, Gaussian-weighted gradient amplitude and tri-linear interpolation are used to obtain the weight. The histogram vectors over the block were then normalized through L2-norm normalization:
υ→υ/(= υ=22+ε2)1/2
where υ indicates a normalized histogram vector over the block, =υ=k indicates k-norm calculation, k equals 1 or 2, and ε is a minimal constant that prevents yielding infinite values. A per-block normalization scheme is intended to compensate for variations of lighting over the input image. All normalized histogram vectors were combined as a full feature vector with size n×m, where n indicates the dimension of the histogram vector over the block, and m indicates the number of blocks in the detection windows(the region of interest). In this paper, m is 45, and n is 36. The combined vectors were then fed to a SVM for object/non-object classification.
However, feature vectors based on HOG are high dimensional. For example, when the bin number was set as 9 and the overlap rate of block set as 0.5, the dimension of the feature vector was 45×4×9=1 620. High dimensional feature brings about the complex and large calculation on feature extraction, training, and classification. As the edge directions of the Zebra stripes are mainly vertical and horizontal, we propose to use the Fisher criterion to remove redundant HOG features. According to the criterion, the feature with the better ability of discrimination shall show larger similarity within a group than that among groups. Through feature ranking, better feature blocks can be selected.
Let ωR be the category of the real currency and ωC be that of the counterfeit, and NR and NC be the number of samples belong to categories ωR and ωC respectively. The within-class scatter Si for ith class, the whole within-class scatter Sw and between-class Sb can be calculated as below:
mi=1/(Ni)∑iXi, i=ωR or ωC(3)
Si=∑i(Xi-mi)(Xi-mi)T, i=ωR or ωC(4)
Sw=∑iSi, i=ωR or ωC(5)
Sb=(mωR-mωC)(mωR-mωC)T(6)
where X denotes the HOG feature extracted from a block, mi denotes the mean feature for class ωR and ωC. The bigger the value of Sb/Sw, the better the discrimination capacity of the feature X.
The HOG feature blocks that have more discrimination information can thus be identified. Given a number of N blocks for HOG feature extraction, the feature block selection process can be described as follows:
a. Calculate the HOG feature Xi for each block.
b. Calculate Fisher score Si(i=1, 2,..., N)for each feature Xi.
c. Sort Fisher score Si(i=1, 2,..., N)in descending order.
d. Select the maximum of Si(i=1, 2,..., N)as feature to input SVM classifier and calculate classification accuracy R.
e. Given preset classification accuracy Rt which meet system requirement, if R>Rt, we stopped adding Fisher score from the rest of Si. Otherwise select a next maximum from the rest of Si(i=1, 2,..., N)to input SVM classifier until the new classification accuracy R is bigger than Rt.
f. Output the selected M HOG feature blocks.
Once the ROI region around the security line was located, the HOG feature was extracted and input to SVM for real and counterfeit currency identification. SVM theory mainly focuses on binary class pattern recognition problems[15]. Let the training set be {(x1, y1),(x2, y2),...,(xn, yn)}, where xi∈Rn and yi∈{-1, 1} represent the HOG feature vectors and the class label, respectively. If the training set can be partitioned by a hyperplane, the hyperplane is expressed as W·X+b=0, where W and b determine the position of the hyperplane. The problem could then be transformed into one on solving the optimal hyperplane to obtain the optimal partition of the training set. Thus, an optimization model function was established:
minω, b, ξ1/2=W=2+C∑ni=1ξi, C≥0,(i=1,2,...,n)
Subject to
yi[(W·Xi)+b]≥1-ξi, ξi≥0(7)
where W indicates the coefficient vector of the separating hyperplane in the feature space, and b indicates the classification threshold. The relaxation factor ξi was introduced given the classification error. C indicates the penalty term of wrongly classified samples, and n the number of training samples.
A decision function was then derived:
sgn(ωTφ(x)+b)=sgn(∑ni=1yiαiK(Xi, X)+b)(8)
where b is equal to yi-∑ni=1yαi(Xi·X), αi indicates the Lagrange multiplier, Xi the known samples, yi the labels of the known samples, X the test samples and K(Xi, X)the kernel function.
The NIR images of the 2005 edition of the 100 yuan RMB banknote were acquired using the NIR light with wavelength of around 850 nm for testing. Our dataset consists of 540 images captured from 500 real notes and 40 counterfeit notes. The position, width and height of the ROI were determined in the preprocessing stage. The sizes of the IR image and ROI are 640×480 and 32×128 pixels, respectively. All the algorithms in this paper were implemented with MATLAB R2011b and they were practical and applicable to similar security features of banknote identification systems.
In this experiment, the positive samples number is 500 and the negative samples number is 40. We randomly split the positive and negative samples into training and testing sets, 50% as training and 50% as testing. While 250 positive and 20 negative samples were used to select HOG features and train SVM, another set of 250 positive and 20 negative samples were used for testing. We repeated the experiment 10 times, and each time used a different 50% of the sample as training and testing.
The accuracy, miss rate, and false positive rate(FPR)were then calculated according to formulas(9)to(11), and compared with those of Ref.[16].
miss rate=(FN)/(TP+FN)(9)
false positive rate=(FP)/(TN+FP)(10)
accuracy=(TN+TP)/(TN+TP+FP+FN)(10)
11)
where TP is the number of true positive instances; FN is the number of false negative instances; FP is the number of false positive instances, and TN is the number of true negative instances.
We first select the most discriminative features using the algorithm presented in section 2. The selection process stopped when 20 HOG features were selected and 99.03% accuracy was achieved. Fig.2 shows the variations of classification accuracy with the number of image blocks selected for feature extraction. One can observe from the figure that the accuracy increase significantly at the beginning, and become stable when the number of block exceeds 20. In the region of interest, these important selected blocks were located at the area of bright and dark edge.
图2 特征块选取数与分类准确率的关系Grid search algorithm was used to determine the optimum parameters(C,γ)of C-SVM by cross-validation. Grid search required less time and had higher cross-validation accuracy. The result of the parameter selection was in Fig.3.
图3 利用网格法选择参数结果Table1 shows the performance of the proposed method in terms of accuracy, miss rate, FPR and efficiency, together with that of approach developed in Ref.[16]. An infrared feature extraction algorithm based on convolution and experience threshold classification method was proposed in Ref.[16]. By using horizontal projection and selecting appropriate convolution kernels, better paper currency identification accuracy is achieved in Ref.[16]. Three methods in table 1 obtain accuracy more than 99%, miss rate no more than 1%, and FPR 0%. However, the method proposed in this paper has higher efficiency. The average detection times for each image for these three methods are 0.85, 0.56 and 0.25 s, respectively. Obviously, the average detection time is greatly shortened when improved HOG descriptors were used. In conclusion, the method in this paper improves efficiency of paper currency identification and achieved better accuracy.
表1 统计结果的比较The IR zebra-stripe anti-counterfeiting pattern is difficult to forge, so recognition of the IR zebra-stripe contributes to RMB banknotes authentication in practical application. In addition, the HOG feature has a better description of shape characteristics of object and is robust for geometric and photometric deformation of image. HOG is selected as feature descriptors in this paper. In order to reduce its dimensionality and improve identification efficiency, an optimized HOG feature was extracted and input to C-SVM for classification, which is used in new application fields. Compared to the state-of-the-art algorithms, the proposed method makes use of infrared imaging and recognizes the forgeries rather than the value. The experiments performed on real notes and counterfeit notes provided by the Chinese bank demonstrate good performance on accuracy and efficiency. Furthermore, this method is robust to low contrast and background noise. Future wok will be devoted to add other security features(e.g. the great hall of the people, watermark, etc)validation to further improve the recognition accuracy. In addition to the Chinese RMB, the extension of current method to other currencies such as US Dollars and HK Dollars will also be considered.
深圳大学学报理工版
JOURNAL OF SHENZHEN UNIVERSITY SCIENCE AND ENGINEERING
(1984年创刊 双月刊)
主 管 深圳大学
主 办 深圳大学
编辑出版 深圳大学学报理工版编辑部
主 编 阮双琛
国内发行 深圳市邮电局
国外发行 中国国际图书贸易集团有限公司(北京399信箱)
地 址 北京东黄城根北街16号
邮 编 100717
电 话 0755-26732266
0755-26538306
Email journal@szu.edu.cn
标准刊号 ISSN 1000-2618
CN 44-1401/N