Zernike Moments and Rotation Invariant Object Recognition. A Neural Network Oriented Case Study
FYSISCH EN ELEKTRONISCH LAB TNO THE HAGUE (NETHERLANDS)
Pagination or Media Count:
This report presents the results of the feasibility study investigating the characteristics of complex Zernike moments and their application in translation-, scale-, and rotation-invariant object recognition problems. The complex Zernike moments are used as characterizing features in a neural network based target recognition approach for the classification of objects in images recorded by sensors mounted on an airborne platform. The complex Zernike moments are a transformation of the image by the projection of the image onto an extended set of orthogonal polynomials. The emphasis of this study is laid on the evaluation of the performances of Zernike moments in relation with the application of neural networks. Therefore, three types of classifiers are evaluated a multi-layer perceptron MLP neural network, a Bayes statistical classifier and a nearest-neighbor classifier. Experiments are based on a set of binary images simulating military vehicles extracted from the natural background. From these experiments the conclusion can be drawn that complex Zernike moments are efficient and effective object characterizing features that are robust under rotation of the object in the image and to a certain extent under varying affine projections of the object onto the image plane.
- Target Direction, Range and Position Finding