Segmentation of fingerprint images on a glass plate : a multi-class fern approach

Details

Supervisors

De Vleeschouwer, Christophe

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en mathématiques appliquées

Abstract

In the scope of a larger work studying the signature electrical signals of particular strains in the fingers, this work addresses the issue of segmenting the contact area between the finger tip and a glass plate, i.e. the fingerprint in the image. The automatized segmentation problem is defined as (Bazen and Gerez, 2001) deciding which parts of the image belong to the foreground, here the contact area, and which part to the background, the noisy area at the borders of the image that can contain various confusing elements for the algorithm like spurious fingerprints, artifacts or light spots. In addition to that, this works adds two more classes, one of each side of the delimitation line between the foreground and the background. This report tries to solve the problem on the provided data set using random ferns, a tool that evaluates the probabilities of each pixel to belong to each one of the classes and allocates the latter to the one that is the most likely, based on a semi-naive Bayesian approach. Each fern consist of several binary tests in the neighbourhood of a pixel based on features of the fingerprint texture. Two of them yield good results: the coherence and the absolute deltas of pixels intensities. After training the ferns probabilities on a small part of the manually segmented images, the results of segmentation are excellent for high-contrasted fingerprints, with more than 90% of pixels correctly classified, a percentage decreasing with the contrast level inside the fingerprint. Lastly, morphological operations are performed on the resulting region to obtain a compact segmented area and reduce the classification errors. Eventually, a few leads for improvement of the method and other potential useful features are discussed in the conclusion