Translation, Scale and Rotation Invariant Features Based on High-Order Autocorrelations

Local high-order autocorrelation features proposed by Otsu have been successfully applied to face recognition and many other pattern recognition problems. These features are invariant under translation, but not invariant under scale and rotation. We construct scale and rotation invariant features from the local high-order autocorrelation features. x 1 Introduction In pattern recognition problem, it is important to extract invariant features from given images. The capability of statistical methods such as discriminant analysis depends on the dimensionality of inputs and the number of samples. If the dimensionality of inputs is large, a large number of samples are needed. Therefore we should construct a low dimensional feature space from a raw image that belongs to an innite dimensional functional space. If some transformation does not change the categorical property of an image, the feature should be invariant under that transformation. The low dimensionality of features is also eective to reduce the computation time of pattern recognition tasks. Otsu proposed a practical simplication of high-order autocorrelation features (high-order local autocorrelation 5)), which is successfully applied to many kinds of pattern recognition problems such as face recognition 6), 3). However, since the high-order autocorrelation features are not invariant under rotation and scale, the performance of recognition decreases under those transformations. In this paper, we construct scale and rotation invariant features (approximately) from Otsu's high-order local autocorrelation features. The diculties of constructing