Region localization based on rotational invariant feature and improved self organized map

The issue of target localization by means of texture analysis is addressed. First the texture feature extraction based on multi-channel Gabor filter decomposition and the rotation invariant representation of Gabor features are analyzed in the view of their ability of classification. After that, a method based on Gabor features and neural network classifier is proposed. The method is composed of two stages, unsupervised texture clustering and target localization. In the first stage, original feature space extracted by Gabor filter banks is applied in training a self organized map classifier and a novel merging scheme is presented to achieve the accuracy of clustering. In the second stage, digital Fourier transform of the original feature vectors are applied in back propagation (BP) network to ensure rotation invariance in localization. In the experiments, the usefulness of the proposed method is demonstrated on texture database and practical barcode localization system as well. The method is also proved rotation invariant and accurate in localizing target texture.