Rotation-invariant MLP classifiers for automatic aerial image recognition

This paper describes the application of Multi Layer Perceptron (MLP) neural networks to the problem of Automatic Aerial Image Recognition (AAIR). The classification of aerial images independent of their orientation is required for automatic tracking and target recognition. Rotation-invariance is achieved by using rotation invariant feature space in conjunction with feed forward neural networks. The performance of the neural network based classifiers in conjunction with 3 types of rotation-invariant AAIR global features: the Zernike moments, central moments, and polar transform are examined. The performance of the Zernike based classifier is compared with that of the classical central moments, and polar transform. The real part of the phase spectrum of the Fourier plane is employed in combination with the MLP for rotation and translation invariance. The advantages of these approaches are discussed. Although a large image data base would be necessary before this approach could be fully validated, the initial results are very promising.