Research On Recognition Of Serial Number In Bank Notes With Complex Background

Research on character recognition including online/offline handwriting recognition and printed character recognition has received a lot attention which plays a more and more important role in daily applications. Many papers have reported promising results in the fields of bank cheque processing, ZIP code and car licence plate recognition. However, little study has focused on the automatic recognition of bank note serial numbers. High accucy bank note serial numbers recognition system have great academic and application values, it will not only facilitate the prevention of forgery crimes, manage the currency circulation, but also have a positive impact on the economy. In this paper, we comprehensively investigate the techniques related to recognition of serial number on RMB (Renminbi) bank notes.First, this paper presents the serial number segmentation and extraction methods including the steps of paper currency scanning, skew correction, orientation identification, serial number region detection and binarizaiton, and serial number extraction. To detect the serial number region, the skew correction is performed by Hough transform and the RMB scanning orientation is identified based on resolution down-samlping and template matching scheme. Instead of extracting the RMB serial number from gray level image, three novel binarization approach using stroke model based feature is applied to the serial number region. We extract the RMB characters by combining apriori knowledge and block contrast average scheme which considers the contrast between foreground and background pixels, and conduct superior performance.Second, the concept of character gray-scale normalization and distortion has been proposed, which can stabilize the generalization performance and improve the classication accuracy. A new database NUST-RMB2013 has been collected from scanned RMB images to evaluate various character feature extraction and classification methods for the recognition of RMB serial numbers. We comprehensively implement and compare two classic and one newly merged feature extraction methods (namely, gradient direction feature, Gabor feature, and CNN trainable feature), and four different types of well-known classifiers (SVM, LDF, MQDF, and CNN). The pros and cons of these methods are discussed according to the experimental results.Third, we propose a novel part-based character recognition method for RMB serial number recognition. Given a training sample, we first generate a set of local image parts using the Difference-of-Gaussians (DoG) keypoint detector. Then, train an SVM classifier with both the original and local image parts. During the test, all the local parts of the input samlpe are classified by an SVM classifier to provide a confidence vector for each part. Three methods are introduced to combine the recognition results of all parts. Since the serial number samples suffer from complex background, occlusion, and degradation, our part-based method takes advantage of both global and local character structure features, and offers an overall increase in robustness and reliability to the entire recognition system.Fourth, five multiple classifier combination strategies (including specially designed linear and cascade combination methods) are presented. Since high reliability is more important than accuracy in financial applications, we introduce three rejection schemes of first rank measurement (FRM), first two ranks measurement (FTRM) and linear discriminant analysis based measurement (LDAM). All the classifiers and classifier combination schemes are combined with different rejection criteria, as a result, nine combined rejection methods are obtained. The experimental results show our cascade method provides an effective way to utilize the complementarities of different classifiers and be able to promise better recognition and rejection abilities than each of the relevant classifier.