Rotation invariant IR object recognition using adaptive kernel subspace projections with a neural network

Abstract

This paper examines two techniques for rotation invariant, adaptive feature extraction and classification of infra red images using a feedforward neural network model. Both approaches use a set of adaptive kernels, or wavelets, to generate rotation invariant features for classification and allow for direct minimisation of a classification error criterion against the input images whilst maintaining a low dimensional parameter space. Each feature extraction parameter is estimated using errors backpropagated from the classification stage.

The first of the two methods uses complex kernels with adaptive radial polynomials. When combined with a magnitude nonlinearity in the first layer of the model they provide rotation invariant features for classification. However, there are several problems with this model which make it impractical. A second method provides a much simpler solution and uses the preprocessing technique of 6 normalisation with a standard adaptive feature extraction and classification model. Both of these methods have been tested on the difficult problem of discriminating between objects derived from a set of real infra red images. Results and discussion are provided in this paper.