Distinctive Point Extraction and Recognition Algorithm for Various Kinds of Euro Banknotes

Counters for the various kinds of banknotes require high-speed distinctive point ex- traction and recognition. In this paper we propose a new point extraction and recognition algo- rithm for Euro banknotes. For distinctive point extraction we use a coordinate data extraction method from specific parts of a banknote representing the same color. To recognize banknotes, we trained 5 neural networks. One is used for inserting direction and the others are used for face value. The algorithm is designed to minimize recognition time by using a minimal amount of recognition data. The simulated results show a high recognition rate and a low training period. The proposed method can be applied to high speed banknote counting machines. Common banknote counting machines only count one single type of banknote. When depositing more then one kind of banknote we first must sort the banknotes based on their face value before counting the total sum. Doing so takes time and is also very complex. To solve these problems, counters for the various kinds of banknotes have been developed. Counting machines for the various types of banknotes require high-speed recognition and counting because the two processes are performed simultaneously. Most recognition algorithms utilize the sizes or colors of banknotes. Gori and Priami (1) and Kim (6) used banknote size and their featured character for recognition. However, it is assumed that the inserted banknote must be authentic. If any paper that has the same size as a banknote is inserted, an error will oc- cur. Furthermore, Kim (6) used a CCD camera to rec- ognize the kind of banknote by applying it to any se- lected area of the image for banknote classification. Takeda and Nishikage (2) used two sensors to in- crease the number of recognition patterns. The pur- pose of the first sensor is discrimination for a known image and the second sensor is for exclusion of an unknown image. But, these methods require too much time to recognize a banknote because the obtained image using a CCD camera is very large and also in- cludes too much information such as noise. Therefore it is unsuitable for high-speed recognition processing. Lee (4) performed training and recognition through neural network and CIS sensor. He did not use the entire image but rather any one selected horizontal line as input data for recognition. It is necessary to reduce the amount of data for high speed recognition. In this paper, to reduce the amount of data required in the recognition process, we proposed a method using a lesser amount of input data than the other methods. For data reduction, particular blocks such as characters of banknotes should be selected. This is considered to be an effective way to reduce the amount of data. We used 4-bit gray scale images of banknotes. There are many black colored parts in gray scale banknote images, particularly the face value number. Black color features are also robust to noise. When noise is added to the black color, the noise is unnoticeable with the exception of some bright color noise. By using this feature, the black colored parts can be a distinctive data of banknotes for recognition and classification. For the banknote recognition process, a back-propagation neural net- work that has input vectors consisting of distinctive points was designed. The input vectors were created from distances between distinctive points and the ori- gin of the unique block. Seven kinds of Euro bank- notes were used as sample banknotes.