Coin Images Seibersdorf - Benchmark

Transcription

1 Coin Images Seibersdorf - Benchmark Michael Nölle 1, Björn Jonsson 2, Michael Rubik 2 ARC Seibersdorf research GmbH 1 Safety & Security, 2 High Performance Image Processing A-2444 Seibersdorf michael.noelle@arcs.ac.at Background The changeover from 12 European currencies to the Euro created a unique situation. Great volumes of money had to be physically returned to the national banks of the member states. Charity organisations took the opportunity to appeal for funds. In Austria alone the charitable donations amounted to several hundred tons of cash. Unfortunately, the coins could only be collected as a potpourri of currencies. The sheer volume of material rules out any attempt to separate the money manually and calls for an automatic processing device. In 2003 a coin sorting device called Dagobert [1.] was built at ARC Seibersdorf research GmbH. The coins which originate from far more than 100 countries were sorted by Dagobert within two years. In total more than 2000 different coin faces of over 600 different coin types had to be trained and comprise the backbone of the recognition unit of Dagobert. The original goal was to sort and recirculate high volumes of coins according to the individual requirements of the national banks. Sorting criteria were thickness, diameter and the images of both sides. In some cases a subsequent separation by further physical characteristics became necessary. The acquired training data comprises a value in itself as it can be used to evaluate object recognition algorithms on a very large set of objects to be recognised. Moreover, an almost unlimited set of test data is available. To foster the development of robust recognition and image search algorithms ARC Seibersdorf research GmbH has defined the Coin Images Seibersdorf (CIS) Benchmark, as a part of the MUSCLE ( ) benchmarking initiative ( Benchmark Preview The initial CIS-Benchmark as presented at the MUSCLE Scientific Meeting in Malaga, 4-5 Nov 2004, contains a small amount of the training data and validated test images with a limited number of coin faces. Presently, there are 1100 test coins (i.e images) and a set of 109 coin types with 389 coin classes available as training data Page 1 / 9

2 Final Benchmark The final CIS-Benchmark contains 693 coin types with 2270 different coin faces as the training object set. In the course of the sorting process images of randomly chosen coins were collected as a test data set. In a first tranche validated test coins (i.e images) are presented. Two additional tranches, validated test coins each, are planned. General Terms The following list defines the terms used below. Coin Type: Coin Class: Front Side: Reverse Side: Test Coin: Validated Image: Coin Class ID / Class ID: Coin ID: Learn/Training Image: Average Image: E.g. Germany 1 DM, USA Quarter, Euro 1 cent Each coin type has at least two coin classes: obverse (front side, showing the value) and reverse side. Example: all Euro coins have one front side (defining one coin class) but country specific reverse sides (defining more than 10 coin classes). The obverse/front side of the coin - showing the value (unambiguously in most cases). The front side defines an F coin class. The reverse side. To recognize a coin type one of the two images has to belong to a front side class and the other one to a reverse side coin class. The reverse side defines an R coin class. Validated coin type, which has to be recognized / mapped to a coin type (2 validated images, one F coin class and one R coin class). Image of one side of a test coin together with the information, which coin class it belongs to. Code for the coin class of the image. Unambiguous number for all test coins to assign the two images to each other. Image of a specific coin class for training purposes. For each coin class there are up to 30 training images. Centred, rotated and averaged image of a coin class. The average image is generated from some of the learn / training images of this coin class. May be used for training purposes Page 2 / 9

3 Coin Types and Coin Classes Each coin type defines at least two coin classes, one for the obverse / front side and one for the reverse side of the coin. The side showing the value of the coin is the front side. Due to variations resulting from the manufacturing process of the coins and/or variations of the coin die there might be more than one coin class for a certain coin type. Each coin class in the CIS-Benchmark is represented by a coin class ID, which is a 9 character string, TTTT-CCCS TTTT gives a four digit coding of the coin type, CCC gives a two digit + one character coding of the coin class, S gives a one digit coding of a coin subclass. The coin type coding starts with a two digit group number (used as a country/currency indicator in most cases), followed by a two digit number group which numbers the possible coin values of this currency group. The numbering of both groups starts with 01. The coin type is separated from the coin class and coin subclass indicator by -. The coin class coding starts with a two digit number group characterising the specific coin class (face) followed by a one character group indicating a front ( F ) or reverse ( R ) coin side. The numbering starts with 01F and 01R respectively. A last digit at the end of each coin ID (starting with 0) marks similar subclasses which need not be distinguished. There maybe coins that are not known within the training set of the CIS-Benchmark. These unknown coin types and coin classes have a special coin class ID Example: R is the coin type Germany 10 Pfennig, 01R specifies a reverse side class, 2 the third sub-class of this coin class, which need not necessarily be distinguished from other coin sub-classes (in this case there are 3 coin sub-classes for this reverse side coin class). Example : is the coin ID for an unknown coin faces Example: F is the coin type Belgium 1 Franc, 03F1 is a front side class which has to be distinguished from 01F1, but not necessarily from the coin sub-class 03F F F F Page 3 / 9

4 CIS-Benchmark data The CIS-Benchmark contains two directories \Training\ \TestData\ which will be described in detail below. All images are stored in png format. Each image has 640 x 576 pixels and consists of two parts: Upper part: image of coin, including data in Line 1 : 640 x 512 pixels Lower part: visible information and logo of Seibersdorf : 640 x 64 pixels Trainings Data The CIS-Benchmark data contains the directory \Training\ with all data sets of known coins and the coin types file. For each coin class (i.e. \Training\02\0204\ R2\ ) there are up to 30 training images. Uncommon coin faces might be represented by only a few or even one training image. The file name of each learn / training image consists of: Class ID (9 characters) + underscore + 2 digits (+.png ) The last two-digit group counts the training image set for this coin ID starting with 00. Additionally, the definition of each coin ID contains an average image (centered and rotated): Class_ID + underscore + AVG.png and a parameter file containing useful information : Class_ID + underscore + params.txt : Content of the parameter file: CoinClass Class_ID (9 characters) Angle_00 angle to rotate the learn image _00 into position of average image. ThicknessMin smallest value of thickness ThicknessMax biggest value of thickness DiameterMin smallest value of diameter DiameterMax biggest value of diameter ImageCount number of images used to calculate the smallest and biggest values. Remark: Thickness and Diameter are measured with limited accuracy as specific light sensors have been used for the measurement! Additionally, the diameter may be calculated from the images themselves. Usually this gives a higher accuracy. The values for DiameterMin and DiameterMax have been calculated directly from the images. The angular position of the average image can slightly differ from the correct upright position. The angle_00 gives the correct angle to rotate the learn image _00 into the rotational position of the average image Page 4 / 9

5 The data is arranged in the following directory structure: Training trainings data directory Group first 2 digits of coin type Type coin type (all 4 digits) Class full 9 character coin class Example : Trainings data for R2 : \Training\02\0204\ R2\ R2_00.png \Training\02\0204\ R2\ R2_01.png \Training\02\0204\ R2\ R2_02.png \Training\02\0204\ R2\ R2_AVG.png \Training\02\0204\ R2\ R2_params.txt Test data The CIS-Benchmark data of tranche 1 contains the directories \TestData\CoinM_00001 \TestData\CoinM_05001 \TestData\CoinM_10001 \TestData\CoinM_15001 \TestData\CoinM_20001 \TestData\CoinM_25001 with data sets of all test coins. Each CoinM_nn001 contains 5000 validated test coins. Each test coin consists of two validated test images from either side of the coin. They are arranged in two different subdirectories SIDE1 and SIDE2. Example: \TestData\CoinM_25001\SIDE1\ \TestData\CoinM_25001\SIDE2\ whereby the two images of one coin have the same coin ID which is used as filename. The 6 digit coin ID is a unique number for all test coins (note: training images are named differently and have no coin ID!), relating to the time of acquisition, but not to any systematic numbering Example: \TestData\CoinM_05001\SIDE1\ png \TestData\CoinM_05001\SIDE2\ png Both directories contain the same number of test images. Which of the two coin sides will be found in SIDE1 is a matter of chance and depends on which side was lying face up when sorted by the machine. The orientation of the second image results from flipping the coin around the vertical axis after taking the first image. This can be used as an additional criterion to determine the correct coin type Page 5 / 9

6 Ground truth The ground truth for each test image has been generated manually and is encoded into the test images. The correct coin class ID of the respective coin side is 1. coded into the first line of image data. The following example shows the hexadecimal and the ASCII coding: E E A A D C I S : 4 5 : 3 7 T R Visible in the lower part of the image (together with diameter and thickness as measured by the corresponding light sensor) 3. Additionally the ground truth is listed in a textfile, \TestData\CoinM_25001\BenchmarkFileList.txt containing a line for every coin_id of the test coins CoinID Vote ClassID_1 Thick1 Diam1 ClassID_2 Thick2 Diam F R F R F R Page 6 / 9

7 Characterization of Information for test and training images in line 1 Same additional information has been encoded into the first line of the images which might be helpful for the recognition process. All the training data were collected on a special training system which had one camera station only (indicated by Source 0 ). The test data was taken from the two sorting systems, which are each equipped with two camera stations. Therefore, each test coin come from either camera 1 and 2 or from camera 3 and 4. For each of the 5 possible camera stations there is a light sensor measuring thickness and diameter with limited accuracy. The following additional information comes along with each image: Coin_ID: unique 6 digit coin identification number (only test images) Date and Time of Acquisition Source : 0 Training system (used only for training) 1, 2 First and second camera position on sorting system 1 3, 4 First and second camera position on sorting system 2 Class_ID (training and test images) Thickness and Diameter as measured by the corresponding light sensor. This data are coded into the filename, the first line of the image, and/or are visible in the lower part of the image. Additionally, they will be presented in a list file. The following table gives an overview where these data may be found. Data Test Image Learn Image Average Image Coin_ID Line1 + filename+list no meaning no meaning Date and Time of Acquisition Line1 Line1 no meaning Source Line1 Line1 no meaning Class_ID Line1 + visible+list Line1+vis.+f.name Line1+visible Thickness and Diameter Line1 + visible+list Line1+vis. no meaning Page 7 / 9

8 Line 1 has following structure: Position Length Data Test I. Learn I. Avg.I. 0 3 CIS (Coin Images Seibersdorf) Yes yes yes 3 1 Version : 1 Yes yes yes 4 1 Source: 0 / 1 / 2 / 3 / 4 yes yes 5 10 Date [ DD.MM.YYYY ] yes yes 15 1 Space 16 8 Time [ hh:mm:ss ] yes yes 24 1 Type of Image T L A 25 9 Class_ID e.g R2 yes yes yes 34 1 Space 35 6 Coin_ID e.g yes 41 1 Space 42 4 Thickness e.g [mm] yes yes 46 1 Space 47 4 Diameter e.g [mm] yes yes Space 81 4 Voting e.g yes 85 2 Space Space Voting: normally equal to coin type (first 4 digits of Class_ID), if one or both sides of a coin are unknown, the whole coin is unknown und receives 0003 as voting result. Coin types file All known coin types (109 in the initial CIS-Benchmark) are listed in the file \Training\CoinTypes.txt which contains one line for every coin type : CoinTyp 4 digit code of the coin type Diam Default diameter [mm] Thick Default thickness [mm] Angle 0 = medal alignment, 180 = coin alignment When a coin is held for vertical viewing with the front side aligned upright and flipped around the vertical axis, the reverse side will be seen upside down for a 180 coin (coin alignment) Page 8 / 9

9 List File The ground truth is also collected in the file \TestData\ CoinM_25001\BenchmarkFileList.txt which contains one line for every test coin: Example: Coin_ID Voting Side1:Class_ID Thickness Diameter Side2: Class_ID Thickness Diameter F R F R R F References [1.] Nölle, Michael and Penz, Harald and Rubik, Michael and Mayer, Konrad and Holländer, Igor and Granec, Reinhard, Dagobert -- a new coin recognition and sorting system, Proceedings of the 7th International Conference on Digital Image Computing - Techniques and Applications (DICTA'03), Macquarie University, Sydney, Australia, December [2.] Fürst, Kronreif, Wögerer, Rubik, Holländer, Penz, Development of a mechatronic device for high speed coin sorting Proceedings of IEEE International Conference on Industrial Technology, Maribor, Slovenia, [3.] Huber, Ramoser, Mayer, Penz, Rubik, Classification of coins using an eigenspace approach, Patt. Rec. Letters, accepted for publ Page 9 / 9