Force field feature extraction for ear biometrics

The overall objective in defining feature space is to reduce the dimensionality of the original pattern space, whilst maintaining discriminatory power for classification. To meet this objective in the context of ear biometrics a new force field transformation treats the image as an array of mutually attracting particles that act as the source of a Gaussian force field. Underlying the force field there is a scalar potential energy field, which in the case of an ear takes the form of a smooth surface that resembles a small mountain with a number of peaks joined by ridges. The peaks correspond to potential energy wells and to extend the analogy the ridges correspond to potential energy channels. Since the transform also turns out to be invertible, and since the surface is otherwise smooth, information theory suggests that much of the information is transferred to these features, thus confirming their efficacy. We previously described how field line feature extraction, using an algorithm similar to gradient descent, exploits the directional properties of the force field to automatically locate these channels and wells, which then form the basis of characteristic ear features. We now show how an analysis of the mechanism of this algorithmic approach leads to a closed analytical description based on the divergence of force direction, which reveals that channels and wells are really manifestations of the same phenomenon. We further show that this new operator, with its own distinct advantages, has a striking similarity to the Marr-Hildreth operator, but with the important difference that it is non-linear. As well as addressing faster implementation, invertibility, and brightness sensitivity, the technique is also validated by performing recognition on a database of ears selected from the XM2VTS face database, and by comparing the results with the more established technique of Principal Components Analysis. This confirms not only that cars do indeed appear to have potential as a biometric, but also that the new approach is well suited to their description, being robust especially in the presence of noise, and having the advantage that the car does not need to be explicitly extracted from the background. (c) 2004 Elsevier Inc. All rights reserved.