Feature Recognition And Discrimination Of RMB Multispectral Image

Over the past decade, banknote image recognition technology is very active in the field of pattern recognition, which has a very important significance for the financial sector, its role is given sorting information such as enominations, towards, being genuine or fake to the control devices.With the promulgation and implementation of the new national standard GB16999-2010’General Technical Requirements for RMB Verifying Instrument’, new requirements such as real-time and correctness is proposed for image recognition. In this paper, a fast banknote sorting method is proposed for the various features of the new version of RMB. And on this basis, the multi-spectral image identification and discrimination is carried out.First, the image brightness compensation is implemented. When the boundary is detected, a very stable algorithm to remove the error points is used, which is suit for variety of defects cases; when tilt is corrected, a method labeled as "rotate when used" is used; when designing recognition algorithm, anti-rotation is priority, which greatly improve time efficiency.In the algorithm of banknotes rapid classification, a banknote image recognition algorithm based on double-sided mesh feature and Multi-angle Gaussian Mixture Model is proposed. In this algorithm double-sided mesh is used for feature extraction, and the number of mesh is computed according to distance between and within classes, which proves the advantages of single-sided mesh compared with double-sided mesh; Then, training different Gaussian mixture model for different angle is to save the time used for entire image rotation process, simplifying the identification process; Improved Gaussian density judgment function is designed to further reduce time-consuming. The experimental results show that the algorithm can get 100% percent recognition rate while recognizing one banknote in 4ms, guaranteeing accuracy and real-time performance.White images, infrared images and ultraviolet images are collected in the multi-spectral image recognition. Algorithms used here include watermark identification based on similarity coefficient, color ink identification based on Euclidean distance, white watermark identification based on template adding up, infrared stripe identification based on low pass filter, Great Hall feature recognition based on infrared contrast and UV feature identification based on highlights adding up. Proceed from a certain kind of image alone, some fake banknotes will be missed. Integrating all means of identification, good recognition results are achieved. All identification can be completed within 50ms.