Pattern Spectra for Texture Segmentation of Gray-Scale Images

This paper presents an unsupervised segmentation of textured images which combines local pattern spectra features and dimensionality reduction techniques. A pattern spectrum is a shape-size descriptor which can detect critical scales in an image and quantify various aspects of its shape-size content. We estimated local features from pattern spectra for discrete graytone images and arbitrary multilevel signals by using a discrete-size family of patterns. Then we applied dimensionality reduction techniques on the features extracted for achieving redundancy reduction and noise reduction. Recently, many neural algorithms have proposed for principal component analysis (PCA) and independent component analysis. In this work, we used two neural PCA and two neural ICA algorithms and compared them.