Banknote Recognition Based On Probabilistic ... - Semantic Scholar

6 downloads 0 Views 362KB Size Report
Banknote Recognition Based On Probabilistic Neural Network Models. COSTAS NASTOULIS1, APOSTOLOS LEROS2, NIKOLAOS BARDIS3. 1Candidate of ...
Proceedings of the 10th WSEAS International Conference on SYSTEMS, Vouliagmeni, Athens, Greece, July 10-12, 2006 (pp802-805)

Banknote Recognition Based On Probabilistic Neural Network Models COSTAS NASTOULIS1, APOSTOLOS LEROS2, NIKOLAOS BARDIS3 3 Adjunct Assistant Professor 1 Department of Automation Candidate of bachelor degree Technological Education Institute of Halkis Department of Automation 34400 Psahna, Halkis, Evia, Greece Technological Education Institute of Halkis 34400 Psahna, Halkis, Evia, Greece 3 [email protected] Research Associate Hellenic Naval Academy 2 Military Inst. of University Education / Hellenic Associate Professor Naval Academy Department of Automation Technological Education Institutes of Halkis 34400 Psahna, Halkis, Evia, Greece Abstract: In the various banking systems around the world there are different types of bank notes whose classification according to their type is a time-consuming process if it is to be done manually by some person. Also such a classification done by a collaborating person usually does not provide the suitable safety required by the banking associations. In this paper we propose a new method for the problem of banknote recognition using a probabilistic neural network (PNN) which is able to proceed to the recognition even if there is an error to the input data in the order of 40%. In addition, the efficiency of the computational resources for the training period and for the recognition rate is compared with some known methods. Key words: - Banknote Recognition, PNN Neural Network

1 INTRODUCTION Because of the different types of banknotes in the banking associations around the world a banknote recognition system had to be created for the classification process which otherwise was timeconsuming to be done manually by a person. Several methods using different techniques have been created as solutions to this problem. Some of them use neural network-based models for the recognition process. In one of these methods the size of the banknote was used as a point for the recognition process, while in another a CCD camera was used to recognize the kind of the banknote. A third method [2] used the approach of sensors-to-

sensors to increase the number of recognition patterns. The first sensor was used to discriminate a known image and the second used for the exclusion of an unknown image. The method in [2], in which our approach is also based on, uses an axis-symmetry mask set for the collection of slab values which are the digitized characteristics of a banknote which in turn are used as inputs to a neural algorithm. In this method a BPN was trained for the recognition process. By using this mask set we are achieving simultaneously recognition of both sides of each banknote. In our approach we used the same mask set for data extraction as in [2] but a different kind of neural network, looking forward to achieve higher recognition rate and a lower training

Proceedings of the 10th WSEAS International Conference on SYSTEMS, Vouliagmeni, Athens, Greece, July 10-12, 2006 (pp802-805)

period. We used a Probabilistic Neural Network which had input vectors consisting of the slab values extracted from the mask sets.

2 DATA EXTRACTION Since we use the slab values as input data for the neural network we must collect several images of the banknotes. For the image collection a scanner was used according to three different conditions which are: i) 72 dpi resolution ii) 256 gray scale level mode iii) 60% scanning image and then a scanned image is saved as a bit map data. To get the NN data the bit map data is divided along the vertical and horizontal axis to get a 32x216 pixels image frame which is the NN data as it is shown in the following picture 1. Equation (1) below is computing the pixel value of the NN pixel data. pixelvalueoftheNNdat a (N p ) =

1 n ∑ xi n i =1

(1)

last raw or column whose gray level is higher than the lowest threshold value are judged to be the edges of the banknote image. When these threshold values have been set appropriately then the edges and the center of the banknote are detected automatically. After detecting the edges and center we are proceeding to the mask set application on the banknote image for the slab value extraction. The mask set that has been applied on the banknote image has 50 mask patterns with each one divided into 48 (4x12) blocks which are used for the location of the mask areas. The following picture 2 shows the configuration of a mask pattern.

Picture 2

In this paper the axis-symmetry mask is applied because of its advantage in recognizing multiple kinds of banknotes. To use this kind of mask, a banknote image and a mask pattern are equally divided pixel axes-wise from the center for both the horizontal and the vertical axis. Also the mask pattern center and the center of the banknote image are located at the same position. After all these processes are completed we are proceeding to the extraction of the slab values which are used as input data in the NN.

Picture 1

Before we locate the masks on the banknotes images, first we must detect the edges and the center of the banknotes. For this purpose we use two significant thresholds values. These threshold values are offsets of the projection of the gray scale level on the vertical and horizontal side of the banknote image, respectively. The first and the

3 RECOGNITION AND CLASSIFICATION The neural network used here is the probabilistic neural network which is suitable for recognition and classification problems. The probabilistic neural networks are a kind of radial basis network and the Matlab command, net = newpnn, creates a new network with a dialog box. That is, the command:

Proceedings of the 10th WSEAS International Conference on SYSTEMS, Vouliagmeni, Athens, Greece, July 10-12, 2006 (pp802-805)

Net = newpnn (P, T, spread) takes two or three arguments, where: P—RxQ matrix with Q input vectors, T—SxQ matrix with Q targets class vectors, Spread—spread of radial basis functions, default value = 0.1, and returns a new probabilistic neural network. If spread is near the zero value the network will act as a nearest neighbor classifier. As the spread value is getting larger, the designed network will take into account several nearby design vectors. The designed probabilistic neural network (newpnn) creates a two-layer network. The first layer has radbas neurons, and calculates its weighted inputs with dist, and its net input with netprod. The second layer has compet neurons, and calculates its weighted input with dotprod and its net inputs with netsum. Only the first layer has biases. The “newpnn” sets the first layer weights to P', and the first layer biases are all set to 0.8326/spread, resulting in radial basis functions that cross 0.5 at weighted inputs of +/- spread. The second layer weights W2 are set to T.

4 SIMULATIONS Using the newpnn Matlab function to create a PNN, then as input vectors are used the slab values extracted from the axis-symmetry mask set process and as corresponded targets are used the numbers corresponding to the value of each euro banknote that are used as simulation banknotes. The probabilistic neural network for the simulations done here (see Picture 3) is a two layer network that compares the input vectors with initial input vectors of the original brand new banknotes that was trained with and classifies the result to the appropriate target class.

Picture 3

Among the seven target classes of the euro banknotes (5€, 10€, 20€, 50€, 100€, 200€, 500€) we created some more target classes for covering the possibilities of a fake banknote or a worn-out (destroyed) one that is no more valid. Also we added in the matrix of the input vectors some more rows that were a combination of the slab values of the banknotes attempting the predictable capability of the PNN neural network for all possible combinations of slab values of a fake or an invalid (worn-out) banknote. After regulating all the process parameters, we were able to proceed to the simulation part. The network was constructed using Matlab code. To the first steps of the code we entered the input vectors and the target classes indices needed for the training period. Then we had to convert the target classes into vectors using the appropriate command. When the training period was completed we proceeded to some tests to find out the recognition ability of our network. We used quite different slab values from the initial ones pretending the use of banknotes having the appropriate deterioration.

5 CONCLUSION The simulation results of the presented studies show that the proposed PNN neural network model was able to recognize and classify with 100% success all the possible versions of banknotes even the fake and the invalid ones. The proposed approach provided close to 100% recognition even if we have 40% alteration to the banknote image in comparison with other

Proceedings of the 10th WSEAS International Conference on SYSTEMS, Vouliagmeni, Athens, Greece, July 10-12, 2006 (pp802-805)

methods [2] which achieved recognition with success less than 80%. The proposed model and method for Banknote Recognition is an effective way of solving the problem for all banking machines accepting even used or worn-out banknotes during the daily transactions.

[4]

References [1] Jac-Kang Lee, Seong-Goo Jeon, and II-Hwan Kim, “Distinctive point extraction and recognition algorithm for various kinds of euro banknotes”, International Journal of control, automation and systems Vol.2, No. 2, June2004. [2] F. Takeda L. and Sakoobunthu H. Satou, “Thai Banknote Recognition Using Neural Network and Continues Learning by DSP Unit”, KES2003 - 7th International Conference on Knowledge Based Intelligent Information and Engineering Systems, 03-05 September 2003, Oxford, UK [3] Nishikage T. and Takeda F., “AxisSymmetrical Masks Optimized by GA for Neuro-currency Recognition and Their Statistical Analysis”, Proceedings of World

[5]

[6]

[7]

Multi-Conference on Systemics, Cybernetics and Informatics, Vol.2, pp.308-314, 1998 Sakoobunthu L., Takeda F., and Sato H., “Thai Banknote Recognition Using Neural Network and Applied to the NeuroRecognition Board”, The International Workshop on Signal Processing Application and Technology, pp.107-114, 2002. Takeda F., Omatu S., and Onami, “Recognition System of US Dollars Using a Neural Network with Random Masks”, Proceedings of the International Joint Conference on Neural Networks, Vol.2, pp. 2033-2036, 1993 Takeda, F. and Omatu S., “High Speed Paper Currency Recognition by Neural Networks”, IEEE Trans. on Neural Networks, Vol.6, No.1, pp.73-77, 1995 Takeda F. and Omatu S., “A Neuro-Paper Currency Recognition Method Using - 54 Optimized Masks by Genetic Algorithm”, Proceedings of IEEE International Conference on Systems, Man and Cybernetics, Vol.5, pp. 4367-4371, 1995

Suggest Documents