Eigenstructure approach to region detection and segmentation

Abstract

A new signal processing method is developed for image region detection and segmentation. The proposed technique formulates the region detection problem in a multidimensional signal processing framework such that a signal structure similar to sensor-array-processing signal presentation is created and the advanced sensor-array-signal processing techniques are employed. Following the region detection, the segmentation is completed by region parameter estimation and pixel classification. The developed method is an unsupervised, non-model-based, eigenstructure approach. It eliminates the ad-hoc assumptions in image modeling, and possesses extensive computation speed superiority over existing model-based approaches. Particularly, it properly utilizes the spatial correlations among the pixels. Details of this technique and several examples are presented. The relationships between this and other methods such as decorrelation, multispectral, and AR model are also discussed.