A Robust Side Invariant Technique of Indian Paper Currency Recognition

A Robust Side Invariant Technique of Indian Paper Currency Recognition

Chetan.B.V

M.Tech Student, Dept. of E&C Engg Malnad College of Engineering Hassan-573201, Karnataka, India

Dr. P.A.Vijaya

Professor and Head, Dept. of E&C Engg Malnad College of Engineering Hassan-573201, Karnataka, India

Abstract

An image based approach which detects Indian paper currencies (Notes) of different denomination has been proposed in this work . This consists of matching an input note image with a database of note images (templates) in two phases. The first phase involves, identifying matching dimension database notes. In the second phase, template matching is performed by correlating the edges of input and matching dimension database note images. The proposed template matching technique provides side invariance and does away with the requirement of placing the front face of the note up. If the correlation coefficient obtained by template matching satisfies the threshold, then input note stands recognised. The entire algorithm is developed in MATLAB and the obtained results are recorded. The proposed technique of paper currency (note) recognition is compared with existing recognition methods like LBP, Image Subtraction, Gabor Wavelet. The comparison results obtained from MATLAB have also been recorded.

1. Introduction

   With development of modern banking services, automatic methods for paper currency recognition become important in many applications such as in automated teller mach ines and automatic goods seller mach ines. The needs for automatic banknote recognition systems encouraged many researchers to develop corresponding robust and reliable techniques [1-8]. Processing speed and recognition accuracy are generally two important targets in such systems. The technology of currency recognition aims to search and e xtract the visible and hidden marks on paper currency for efficient classification. Until now, there are many methods proposed for paper currency recognition. The simp lest way is to make use of the visible features of

   the paper currency, for exa mp le, the size and texture of the paper currency [1]. However, this kind of methods has great limitat ions as currencies of different values may have the same size in some countries, and the visible ma rks may be contaminated by noise.

   Junfang Guo et a l. [1] proposed a method for paper currency based on the traditional local binary pattern method for feature e xt raction and template matching is performed in this method. Neural networks (NN) are wide ly used in the fie ld of paper currency recognition. Fumia ki Ta keda et al. [2] proposed a paper currency recognition method using neural networks. Two types of data sets, time series data and Fourier power spectra, are used. Er-Hu Zhang et al. [3] presented a method using linear transform of gray image and carried out sorting recognition by three layers BP NN. Sigeru Omatu et al. [4] proposed a local principal co mponent analysis (PCA) method, which is applied to remove non-linear dependencies among variables and e xtract the main principal features of data. Fumiaki Takeda et al. [5] proposed a Neuro-Paper Currency recognition method using optimized masks by Genetic Algorith m. Trupti Pathrabe et al. [6] proposed a paper currency recognition system using characteristics extraction and negatively correlated NN ensemble. Three characteristics of paper currencies are considered here including size, colour and texture . Pa rminder Singh Reel et al. [7] proposed a paper currency recognition method, involving the heuristic analysis of characters and digits of serial nu mber of Indian currency notes to recognition of currency notes. CAO Bu-Qing et a l. [8] proposed a method based on uncertain network model structure and indeterminate in itia l we ights and slow convergence speed for Back Propagation Neural Networks, then form GA-BP model was applied for currency recognition.

   Although the NN technology has the ability of self – organization, genera lizat ion and paralle l processing, and has been a good fit for pattern recognition, it also has some wea kness. First, it needs a large number of

   training samp les, which a re used to avoid over-fitt ing and poor generalization. Second, if the distribution of training samp le is not uniform, the result will probably converge to a local optima l or will even diverge unreasonably.

   In currency circulat ion, the original informat ion on

   paper currency may have a loss because paper currency may be worn, blurry, or even da maged. Figure 1 illustrates this problem. Figure 1 shows three notes of five rupee Indian currency. Te xture of each note is diffe rent fro m other. The above image based methods [1-8] of paper currency recognition would not yield proper result in this case, as the intensity values change fro m image to image of same denomination/pattern paper currency.

   To counter this problem the proposed method makes

   use of edge detection and correlating the edge detected images during template matching. After edge detection the image matrix consists of 1s or 0s, instead of 0-255 values. Hence, the proposed method becomes independent of the intensity values of each pixe ls of note (paper currency) image. This gives better recognition results for the proposed method.

   Figure 1. Three paper currencies of same de nomi nation

2. Assumptions

   Following parameters are kept constant during image acquisition:

   Lighting condition

   Distance and Position Perpendicular image acquisit ion

   Also, care is taken so that the surface of the paper currency is clean.

3. Proposed Side Invariant Paper Currency Recognition Method

   Database creation phase

   Image Segmentation

   Image Acquisition

   Saving the Resultant as a Test Image in Database

   Template matching

   Dimension matching

   Object Image Segmentation

   Object Image Acquisition

   Recognition result

   Coin recognition phase

   Figure 2. Bl ock diagram of the rec ogniti on process

   Fig. 2 shows the block diagram of overa ll recognition process. First a database of coin images is created and then recognition is carried out. The different stages involved in the overall process are described one by one.

   1. Image Acquisition and Segmentation

      A digital ca mera is used for image acquisition. Since each note has two faces (sides), there are four possible ways of placing a note for image acquisition. These four possible ways are shown in figure 3. The proposed method is side invariant and hence, is capable of recognising the paper currency in all four cases shown in figure 3.

      Figure 3. Four possible ways of placing a note for image ac quisition

      The next step of the paper currency recognition system would be image segmentation, i.e . separating the note image fro m the background. In first step of segmentation, the edges of input image are detected, ma king use of sobel operator. In second step, filtering of noisy edges other notes edges is performed. In third step, the boundary coordinates of the note part in the input image are noted down. With the help o f the

      boundary coordinates the note part in cut out of the original input image. Thus, the note is segmented for future processes. Figure 4 shows an exa mple of segmentation, employed in th is work.

      Figure 4. Paper currenc y segme ntation result

   2. Dimension Matching

      After paper currency segmentation, the numbers of pixe ls row-wise and number of pixels colu mn-wise in the segmented image are noted. Thse pixe l counts (row-wise and column-wise) give the dimensions of the paper currency in terms of pixe ls.

      After finding the dimensions of the input note, its dimensions are matched with the dimensions of all database notes. The matching dimension database notes are noted down. Hence, during te mplate matching only these matching dimension database notes are matched with input note for recognition of input note.

   3. Template Matching

      The matching is performed by correlat ing the edges of input and database notes. The correlation coefficient is given by,

      Figure 5. Fl ow chart of the te mplate matc hing algorithm

      1. Edge De tection. It is the preliminary step of template matching. Edge detection of the input note and database note, which is about to be matched is performed. Figure 6 shows an exa mp le result of edge detection method which is e mployed in th is work. The edges of input and database notes are correlated to obtain a correlation coeffic ient. This coefficient is the matching score obtained by matching input and database notes.

         r m n

         m n

         Amn

         Amn

         2

         A

         A Bmn

         m n

         B

         2

         Bmn B

         (1)

         Figure 6. Edge detec tion result

         where, r is correlat ion coefficient, A & B a re two image matrices, m & n are nu mber of rows & colu mns

         respectively in both A & B, A & B are the means of matrices A & B respectively.

         New image is selected based on the dimension matching from the database

         Equation 1 is used to obtain the matching score (MS) of the template matching process. Figure 5 shows the entire process of template matching.

      2. Te mpl ate matching by displ ace ment of database note. To obtain accurate matching results, displacement of database note is performed during matching process. Each matching dimension database note is displaced in two shifts. In first shift, the database note is displaced side-wards in terms of one pixe l at each step, from 1 to 5 p ixe ls. At each step, database note edges are correlated with edges of input note. The ma ximu m corre lation coefficient out of these steps is noted. In second shift, the database note is displaced downwards in terms of one pixe l at each step,

         Template matching between input and test image (by displacing the test image side-wards and downwards in terms of one pixel at each step), and correlation coefficient is recorded.

         www.ijert.org 3

         fro m 1 to 5 pixe ls. The second shift is performed at the first shift position where ma ximu m correlat ion coeffic ient is recorded. The ma ximu m correlat ion coeffic ient obtained after second shift becomes the final matching score between input note and particular database note.

      3. Threshol d c omparison. The matching score obtained from each matching dimension database note is compared with a predefined threshold value. If the matching score is greater than threshold, the particular database note stands matched with the input note. Otherwise, the particular database note stands unmatched.

   4. Decision Making

      During threshold comparison, there is a chance of more than one note yielding matching score greater than threshold. So, there is a need of decision making. The database note yielding the ma ximu m matching score is taken as final match of the input note. From the final match database note the denominat ion of input note is concluded. Hence, the input note stands recognized.

4. Experime nts and Results

   To test the accuracy of recognition of the proposed paper currency recognition method various experiments have been performed. The results of various e xperiments have been recorded in this section.

   1. Recognition Results

      1. Recogniti on of noisy (c ontami nate d) paper currencies (notes). The proposed method works well even if the surfaces of the notes are contaminated by noise. This contamination is due to frequent usage of notes as shown figure 1. To prove this capacity of the proposed method, a 5 rupee Indian note (contaminated) has been placed in all possible ways as shown in figure 3 and subjected to recognition test. The recognition results of all cases are shown in figure 7.

         Figure 7 shows that the proposed method effectively

         recognizes contaminated notes. Similar tests can be done by adding Gaussian and salt & pepper noises to input note by means of simulation. A 5 rupee Indian note added with simulated noise has been subjected to recognition tests.

         Figure 7. Rec ognition results of contaminate d notes of all four cases

         Figure 8. Rec ognition result of input note wi th gaussian noise (me an = 0 and vari ance = 0.05)

         Figure 9. Rec ognition result of input note wi th salt and pe pper noise (noise density = 0.1)

         Figure 8 shows the recognition result of input note with

         Gaussian noise and figure 9 shows the recognition result of input note with salt and pepper noise.

      2. Recogni tion capacity of the pr oposed method. To determine the recognition capacity of the proposed method an e xperiment is perfo rmed. Two sets of notes (paper currencies) containing 28 patterns have been

         formed. Pattern is nothing but one of the two faces/sides of a note. Each set consists of a mix of new and old/contaminated notes having denominations five, ten, twenty, fifty, hundred, five hundred and thousand rupees. Each note of one set is matched with all matching dimension notes of the other set. Therefore, a note of one set has to be matched with same pattern of other set and also with different patterns of same dimension. Hence matching and non-matching scores are obtained for each note. By obtaining these scores for all notes, a recognition capacity has been plotted, as shown in figure 10.

         Figure 10. Recogni tion capacity gr aph of the pr opose d method

         Figure 10 shows that the scores of different pattern curve (red) overlaps only a sma ll portion of the scores of same pattern curve (b lue-discontinuous curve). This means that the proposed method has good recognition capacity. It can successfully reject different patterns and accept simila r patterns. From this graph the threshold value can also be decided, which can be used in threshold comparison step. The following para meters can be obtained from this graph:

         Threshold: 0.040105

         True Acceptance Ratio (TAR): 0.928571 True Re jection Ratio (TRR): 0.967213 False Acceptance Ratio (FA R): 0.032787 False Re jection Ratio (FRR): 0.071429

         Here, TAR and TRR g ive the measure of recognition accuracy of the proposed method. FAR and FRR give the measure of error. By observing these values, it can be concluded that the proposed method gives more than 90% recognition accuracy.

   2. Comparison with other existing paper currency recognition methods

      Four input sets have been formed, each containing 28 patterns (new/old/contaminated) for testing the recognition accuracy of the proposed method. Th ese four sets of notes have also been subjected to recognition by means of Local Binary Pattern (LBP) based method, Image Subtraction method and Gabor wavelet based method. The results have been tabulated in table 1.

      Table 1. Comparison of accur ac y of propose d method with other methods

      Method	Accuracy obtained for In put sets (in terms of percentage)				Overall accuracy
      	Input set 1	Input set 2	Input set 3	Input set 4
      Proposed	100	100	99	98	99.5%
      Gabor wavelet	65	70	50	75	65%
      Image Subtraction	60	40	50	55	51.25%
      LBP	60	55	50	45	52.5%

      Fro m the above table it can be concluded that the proposed method is better than other recognition methods.

      Further, fo r much better co mparison, the recognition capacity graphs have been plotted for all three methods. Sa me procedure is followed to obtain them, as discussed while plotting the recognition capacity graph of proposed method.

      Figure 11. Recogni tion capacity gr aph for Gabor wavelet base d me thod

      Figure 11 shows that there is about 75% accuracy of recognition and about 25% error, in Gabor wavelet based recognition method. Figure 12 shows the

      recognition graph of Image Subtraction method. Figure

      12 shows that there is about 60% accuracy of recognition and about 40% of recognition error. Figure 13 shows the recognition capacity graph of LBP based recognition method. Figure 13 shows that there is about 50% recognition accuracy only and about 50% of e rror.

      Figure 12. Recogni tion capacity gr aph for Image Subtrac tion method

      Figure 13. Recogni tion capacity gr aph for LBP base d method

      Fro m figures 10, 11, 12 and 13, it can be concluded that the proposed method has better recognition capacity than other methods in consideration.

5. Conclusion

The proposed method for paper currency recognition has been found to be s imple and accurate. This method yields an accuracy of more than 90%, which is much greater than accuracies of other methods. This method provides side invariance for recognition process. The method also avoids the dependence on constant light

factor during image acquisition up to certain extent. Also two subsequent phases, dimension matching and template matching have been provided to give precise results.

Future works will include modifications of the present technique to recognize paper currencies. The work will include me rging of other image processing techniques, such as, neural networks training using edge detection which would co mplete ly e xtricate the process from the dependency over standard light intensity and standard distance between the note and the camera during image acquisition, adding on to the accuracy of the process.

Acknowledgme nts

Thanks to Mr. Hiren Parmar, Director, VED Labs, Bengaluru, for suggestions and support.

References

1. Junfang Guo, Yanyun Zhao, Anni Cai, A Reliable method for paper currency Recognition based on LBP, Proceedings of IC-NIDC, 24-26 Sept. 2010, pp 359-363.

2. Fumiaki Takeda and Sigeru Omatu, High speed paper currency recognition by neural networks, IEEE Transaction on Neural Networks, Vol.6, Iss. 1, Jan 1995, pp 73-77.

3. Er-Hu Zhang, Bo Jiang, Jing-Hong Duan and Zheng- Zhong Bian, Research on paper currency recognition by neural networks, Proceedings of the Second International Conference on M achine Learning and Cybernetics, Xian, Vol. 4, 2-5 November 2003, p p 2193

   2197.

4. Sigeru Omatu, M ichifumi Yoshioka and Yoshihisa Kosaka, Bank Note Classification Using Neural Networks, ETFA, 25-28 Sept. 2007, pp 413-417.

5. Fumiaki Takeda and Sigeru Omatu, A Neuro-Paper

   Currency Recognition M ethod Using Optimized M asks by Genetic Algorithm, IEEE International Conference on Systems, M an and Cybernetics, Vol. 5, 22-25 Oct 1995, pp 4367 4371.

6. Trupti Pathrabe and Dr. N. G. Bawane, Paper Currency

   Recognition System using Characteristics Extraction and Negatively Correlated NN Ensemble, International Journal of Latest Trends in Computing, Volume 1, Issue 2, December 2010, pp 121-124.

7. Parminder Singh Reel, Gopal Krishan and Smarti

   Kotwal, Image Processing based Heuristic Analysis for Enhanced Currency Recognition, International Journal of Advancements in Technology, Vol 2, Issue 1, January 2011, pp 82-89.

8. CAO Bu-Qing and LIU Jian-Xun, Currency

   Recognition M odeling Research Based on BP Neural Network Improved by Gene Algorithm, Proceedings of Second International Conference on Computer M odeling and Simulation, Vol. 2, 22-24 Jan. 2010, pp 246 250.

9. Frank Y. Shih, Image Processing and Pattern Recognition Fundamentals and Techniques, John Wiley and Sons Publications, IEEE Press, 2010.

10. John C. Russ, The Image Processing Handbook, 3rd Edition, CRC Press, 1998.

11. Anil K. Jain, Fundamentals of Digital Image

    Processing, Prentice Hall of India.

12. Chris Solomon and Toby Breckon, Fundamentals of Digital Image Processing A Practical Approach with Examples in M ATLAB, John Wiley & Sons Ltd, 2011.

Authors Profile

Chetan.B.V obtained his B.E Degree (Electronics and Co mmunicat ion) in July-2010, fro m SJM Institute of Technology, Chitradurga, Karanataka, under Visvesvaraya Technological Un iversity Be lgaum. He is currently pursuing his M.Tech (Digita l Electronics and Commun ication Systems ) Degree since Septe mber-2010, in Malnad College of Engineering, Hassan, Karnataka, India.

Dr. P. A. Vijaya obtained her B.E. fro m M CE, Hassan in 1985, M .E. fro m IISc, Bangalore in 1991 and Ph.D. fro m IISc, Bangalore in 2005. She is currently serving as Professor and HOD, in E&C Depart ment, MCE, Hassan, Karnataka, India. Her a reas of interest are Pattern Recognition, Data Mining, Image processing, Operating Systems, Co mputer Architecture, Microprocessors, Embedded Systems and Real time systems. She has 25

national/international publications to her credit.