Share Email Print
cover

Proceedings Paper

Image classification independent of orientation and scale
Author(s): Henri H. Arsenault; Sebastien Parent; Sylvain Moisan
Format Member Price Non-Member Price
PDF $14.40 $18.00

Paper Abstract

The recognition of targets independently of orientation has become fairly well developed in recent years for in-plane rotation. The out-of-plane rotation problem is much less advanced. When both out-of-plane rotations and changes of scale are present, the problem becomes very difficult. In this paper we describe our research on the combined out-of- plane rotation problem and the scale invariance problem. The rotations were limited to rotations about an axis perpendicular to the line of sight. The objects to be classified were three kinds of military vehicles. The inputs used were infrared imagery and photographs. We used a variation of a method proposed by Neiberg and Casasent, where a neural network is trained with a subset of the database and a minimum distances from lines in feature space are used for classification instead of nearest neighbors. Each line in the feature space corresponds to one class of objects, and points on one line correspond to different orientations of the same target. We found that the training samples needed to be closer for some orientations than for others, and that the most difficult orientations are where the target is head-on to the observer. By means of some additional training of the neural network, we were able to achieve 100% correct classification for 360 degree rotation and a range of scales over a factor of five.

Paper Details

Date Published: 1 April 1998
PDF: 6 pages
Proc. SPIE 3402, Optical Information Science and Technology (OIST97): Optical Memory and Neural Networks, (1 April 1998); doi: 10.1117/12.304959
Show Author Affiliations
Henri H. Arsenault, COPL/Univ. Laval (Canada)
Sebastien Parent, COPL/Univ. Laval (Canada)
Sylvain Moisan, COPL/Univ. Laval (Canada)


Published in SPIE Proceedings Vol. 3402:
Optical Information Science and Technology (OIST97): Optical Memory and Neural Networks
Andrei L. Mikaelian, Editor(s)

© SPIE. Terms of Use
Back to Top