My Bachelor Thesis 

Handrec:
Biometric recognition of persons through hand geometry measurements

Only geometrical features as the length and width of the fingers are compared in order to identify a person.

Abstract
In a lot of situations of our everyday life one has to prove its identity. This happens for example with access controls at buildings, border controls or check in at airports, money automats, on the PC, as well as with e-Banking.

Here the biometric identification and verification takes place and distincts a person by means of its unique, individual characteristics e.g. face, finger mark, iris or hand geometry identifiers.

The verification with hand geometry mesurements compares a scan of a hand with with reference images stored in a database.

Task
Different existing solutions for hand recognition had to be evaluated in a bibliographic research. Then one (or a combination) of these techniques had to be implemented (in Java).

Hand geometry
Hand geometry means the geometrical features of a human hand: The features are the width of the fingers at several positions, the length of the fingers, the width of the palm or the shape of fingers and palm.
Features as skin color or texture of the hand are not used.

Transformations an manipulations
The original image has to be transformed several times in order to get an abstract view and to be able to measure widths and lengths.

work with image
read color BMP image
grayscale filter
threshold filter (binary image)
detect edges (only contour of the hand)
follow border with a modified Pavlidis algorithm (the points are stored in a vector)

work with stored vector
find extremas (fingerpoints and valleys between the fingers)
get three additional points in order to be able to extract the parameters
extract parameters (lengths and widths of fingers)
store parameters as a vector in an external file
compare two vectors
decide if the two vectors represent the same person

See the screenshot images here.

The image has to be scaled for the output in the GUI. The manipulations with the data from the vector are displayed on the image in the GUI.

Test results
The images were taken from the Biomet database. The pictures were taken from 91 persons during 3 different sessions. In general there were 5 pictures per person (3 + 1 + 1). The database consisted of 421 hand pictures.

The hands were NOT fixed with pegs to standardize the scan. They were freely positioned on a standard scanner!

The best results, as expected, were achieved with 5 features per finger (length + 4 widths).

Test set 1:

correct accepting rate of 79,41% by a false accepting rate of 1,47%
correct accepting rate of 85,30% by a false accepting rate of 2,94%

Test set 2:

correct accepting rate of 72,46% by a false accepting rate of 1,45%
correct accepting rate of 75,36% by a false accepting rate of 2,90%

 The results are not extraordinary but nevertheless comparable with other groups.

Screenshot from Handrec

Documentation