Laurens van der Maaten MICC-IKAT, Maastricht University

unimaas.nl

Laurens van der Maaten MICC-IKAT, Maastricht University

Fast and reliable coin

classification

Laurens van der Maaten

MICC-IKAT, Maastricht University


Introduction


Overview

The challenge

System overview

Results

Conclusions


The challenge

Coins (300 tons): 698 coin classes with 2,070 different faces

Classify 5,000 coins (= 10,000 coin images)

Within 8 hours on desktop machine

At least 70% correct

Misclassifications very expensive

Classification as unknown is allowed


The challenge


System overview

Four main components

Segmentation

Feature extraction

Classification

Verification


Segmentation

The extraction of the coin from its background

at the border of a coin, there is an edge

the magnitude of the derivative of the 2D image function

is relatively high at an edge

we can estimate the derivative and threshold it

this is called edge detection


Segmentation

The result of the segmentation:

In addition, we check whether the segmentation is successful


Feature extraction

Measure statistical properties of the coin image that are

typical for this coin type

The statistical properties allow us to discriminate between

coins

The discriminating feature of a coin is its stamp


Feature extraction

Measure statistical distributions of stamp pixels over coin

stamp pixels can be found by edge detection

coin border edges are ignored


Feature extraction

Edge distance distributions

estimate distribution of

distances of edge pixels

to the center of the coin

rotation invariant

measured on multiple

scales


Feature extraction

Edge angle distributions

estimate distribution of

angle of edge pixels

w.r.t. the baseline

rotation invariance

obtained by computing

magnitude of Fourier

transform


Feature extraction

Edge angle-distance distr.

incorporate both

angular and distance

information

measured using 2, 4, 8,

and 16 distance bins

and 180 angle bins


Classification

The feature vector can be considered as a point in a high-

dimensional space

Points that are close together in the high-dimensional space,

are likely to have the same class

Classifiers employ this property to construct decision lines

(i.e. lines that discriminate between coin types)


Classification

Our classifier is trained on a trainingset of around 20,000

coin images

When an unlabelled coin image is input into the system, the

classifier predicts the most likely class of the coin image

The classifications of both coin sides are combined


Verification

The classifier does not provide insight in the probability that

its classification is true

Because errors are expensive, misclassifications have to be

eliminated

This is done by a direct comparison of the coin images in

order to judge their similarity


Verification

The coin is compared with the prototype corresponding to the

classification

This is done for all rotations of the coin


Results

The system classifiers 81% of the coins correctly, while

misclassifying only 0.3% of the coins

The system uses approximately 2 seconds of computation

time per coin image


Conclusions

The system allows for automatic sorting of large amounts of

European coins

Future work: further reducing the amount of

misclassifications

Future work: classification of historical coins


Thanks

Questions?

More magazines by this user
Similar magazines